Skip to content

feat(steer)!: learn of later events by listing the platform's own run records - #7451

Closed
waynesun09 wants to merge 6 commits into
steer-provenancefrom
steer-monitor
Closed

waynesun09 wants to merge 6 commits into
steer-provenancefrom
steer-monitor

Conversation

@waynesun09

@waynesun09 waynesun09 commented Sep 18, 2026 •

Copy link
Copy Markdown
Member

A run already working on a work item now learns when that item changes, and absorbs the update
into the session it is already running.

It learns by listing the execution platform's own run records for its own shim: every legitimate
update already fires the repository's shim, so those records are a server-side log of what
happened since this run started. The listing uses the job token the stage already holds, is
filtered server-side to one workflow file and to runs created at or after this run's start, and
is re-derived from that start on every poll — there is no cursor, no input driver, and no
scheduling inside fullsend run. A poll happens on an interval and immediately at every turn end,
which is the moment a steer is free: the run either absorbs something and takes another turn, or
settles.

max_steers and poll_interval_seconds become live here; until now nothing read them.

A pull request's labels count as an update, as an issue's do. A label added mid-run, such as the
ready-for-review the coder App adds a second after opening a pull request, reaches the agent as
state context instead of leaving that follow-up with an empty delta that is judged but never
delivered.

What this changes for a repository

Two things are visible without opting in to anything:

  • Every Actions run in the repository is retitled <owner>/<repo>#<number>. The per-repo shim
    needs a run-name because issue_comment and issues runs carry no pull_requests[], and
    display_title is then the only server-side field an in-flight run can bind a follow-up to.
    Dashboards, saved filters and scripts that match runs by title need updating.
  • fullsend steer <url> "<text>" reaches a run in flight from a workstation. It posts the
    stage's own slash command rather than a command of its own — the existing route arms already
    carry the right authorization floor for each stage, and the watcher accepts a follow-up run on
    provenance rather than on which words the comment opened with.

Not here yet: the queued run still redoes the work

This PR delivers steers and settles. It writes no processing receipt, so the run queued behind
an absorbed update cannot tell the work is done and does it again from scratch. Until the receipt
lands, absorbing an update therefore costs more than cancelling would: the run in flight reviews
the new head, and the queued run reviews it again.

That is the reason steering stays off by default here (DefaultSteerEnabled is unchanged), and
the receipt is the next decision in the stack rather than an optimization to follow later.

Declining is a real outcome

A run declines steering, and says so, when its environment cannot support it — no job token, no
run id, an ambiguous stage job, a forge login it cannot resolve, or a pi model chain that can fall
back (the runtime would refuse the steer, so the run says so before its watcher starts). An
unresolvable login is not a degradation to a shorter exclusion list: the logins a run posts under are what keep it from
steering itself on its own output, and the post-fix and post-code comments carry no marker to
recognise them by, so no steering is the correct answer.

FULLSEND_STEER_ACTIVE follows the same rule in the other direction. The agent definitions treat
the runner-update opening line as an injection attempt unless that variable is set, and the line
is public, so it is exported only once an iteration's watcher is actually running — never at
bootstrap, where the watcher's live API calls have not happened yet, and never inherited across
validation-loop iterations.

Base and scope

Based on #7447, which decides where an absorbed update's authority comes from. This PR decides
only how a run finds out there is one, recorded in ADR 0119. The six commits are the watcher loop
and the run listing; the runner wiring and the eligibility rules; the run-name binding and the
CLI; the docs; declining a pi run that can fall back before its watcher starts; and reporting a
pull request's label changes.

Implements the design recorded in #6957 (closed as completed; this stack ships it).

Verification

go build ./..., go vet ./... (including test packages), -race on internal/steerwatch and
internal/cli, plus internal/scaffold and internal/forge/..., and pre-commit across the
branch. One unrelated failure reproduces on the base without these commits:
TestApplyCIServerTLSCA_TrustsPrivateCAAndRejectsUntrusted depends on the host's system
certificate pool.

@waynesun09
waynesun09 requested a review from a team September 18, 2026 14:50
@waynesun09 waynesun09 added the fullsend-no-fix Skip bot-triggered fix agent runs label Sep 18, 2026
@qodo-code-review

Copy link
Copy Markdown

PR Summary by Qodo

Steer active runs using GitHub Actions follow-up records

✨ Enhancement 🧪 Tests 📝 Documentation ⚙️ Configuration changes 🕐 40+ Minutes

Grey Divider

AI Description

• Poll GitHub Actions records to discover authorized updates and steer active sessions.
• Add workstation steering and work-item run titles for reliable follow-up binding.
• Bound steering by run-wide caps, token lifetime, deadlines, and explicit settlement.
Diagram

graph TD
  U["User or event"] --> S["Repository shim"] --> R[("Run records")] --> W["Steer watcher"] --> D["Work item delta"] --> A["Agent session"]
  S --> Q["Pending run"]
  A -->|turn end| W
  W -->|settle| A
Loading
High-Level Assessment

The following are alternative approaches to this PR:

1. Poll forge event streams
  • ➕ Could support multiple forge platforms through normalized event drivers.
  • ➕ Provides direct access to source events rather than inferred workflow activity.
  • ➖ Requires race-safe cursors, normalization, and per-forge drivers.
  • ➖ Duplicates authorization and routing already performed by the repository shim.
2. Deploy a courier or relay
  • ➕ Could push updates with lower discovery latency.
  • ➕ Centralizes delivery observability and retry behavior.
  • ➖ Adds another deployed, credentialed, highly available component.
  • ➖ Reimplements routing predicates and expands the trusted infrastructure boundary.
3. Let agents poll at turn end
  • ➕ Requires little runner-side orchestration.
  • ➕ Avoids workflow-run listing costs during active turns.
  • ➖ Depends on model compliance and only discovers changes at turn completion.
  • ➖ Cannot deterministically inject updates during a long-running turn.

Recommendation: Keep the PR's run-record polling approach. It reuses an authenticated, server-side artifact produced by the existing dispatch path, preserves the queued run as an at-least-once fallback, and avoids adding cursors, event drivers, or infrastructure. The alternatives remain useful future options only if platform coverage or polling cost becomes the dominant constraint.

Files changed (31) +3472 / -24

Enhancement (10) +1252 / -13
root.goRegister the steer subcommand +1/-0

Register the steer subcommand

• Adds 'fullsend steer' to the root Cobra command.

internal/cli/root.go

run.goIntegrate steering into the agent run lifecycle +163/-13

Integrate steering into the agent run lifecycle

• Captures steering credentials and identity, starts one watcher per validation iteration, forwards turn ends, and preserves seen runs, baselines, and steer counts across iterations. It also aligns sandbox deadlines, heartbeats, runtime contexts, and timeout reporting with the token-bounded run budget.

internal/cli/run.go

steer.goImplement runner-side steering orchestration +475/-0

Implement runner-side steering orchestration

• Adds eligibility and identity checks, deadline calculations, watcher construction, synchronized runtime delivery and settlement, and turn-end signaling. Steering fails closed when required GitHub Actions identity or authorization context is unavailable.

internal/cli/steer.go

steercmd.goImplement the fullsend steer command +217/-0

Implement the fullsend steer command

• Adds GitHub work-item URL parsing and posts an existing stage slash command as the authenticated user. The forge resolves whether the target is an issue or pull request before selecting or validating the stage.

internal/cli/steercmd.go

fake.goSupport time-filtered workflow runs in the fake forge +24/-0

Support time-filtered workflow runs in the fake forge

• Implements 'ListWorkflowRunsSince' over fake workflow fixtures with timestamp filtering and injected-error support.

internal/forge/fake.go

forge.goExpose time-filtered workflow-run listings +3/-0

Expose time-filtered workflow-run listings

• Extends the forge client contract with workflow-specific run listing from a supplied creation time.

internal/forge/forge.go

github.goList GitHub workflow runs since a timestamp +55/-0

List GitHub workflow runs since a timestamp

• Adds server-side workflow and creation-time filtering with provenance conversion and bounded pagination. The page cap controls the watcher's recurring Actions API cost.

internal/forge/github/github.go

ci.goDeclare GitLab run-record steering unsupported +7/-0

Declare GitLab run-record steering unsupported

• Implements the expanded forge interface by returning 'ErrNotSupported', matching the current GitHub-only watcher gate.

internal/forge/gitlab/ci.go

actions.goAdd chronological follow-up run discovery +27/-0

Add chronological follow-up run discovery

• Extends the watcher's Actions interface with time-filtered listings and sorts discovered runs oldest first before evaluation.

internal/steerwatch/actions.go

watcher.goImplement the poll, steer, and settle loop +280/-0

Implement the poll, steer, and settle loop

• Polls on intervals and turn ends, validates unseen candidates, folds accepted runs into one state-based steer, and tracks consumed state. It settles exactly once on inactivity, cancellation, deadline, low remaining time, or exhausted steer budget.

internal/steerwatch/watcher.go

Tests (9) +1899 / -9
github_client_test.goAccount for steering token reads in auth checks +7/-4

Account for steering token reads in auth checks

• Updates the credential-read allowlist explanation for the job and role tokens intentionally captured by the runner.

internal/cli/github_client_test.go

run_test.goUpdate iteration environment tests for steering state +5/-5

Update iteration environment tests for steering state

• Adapts existing environment command and write tests to the new steering-activity argument while preserving unsteered behavior.

internal/cli/run_test.go

steer_test.goTest steering eligibility and lifecycle integration +728/-0

Test steering eligibility and lifecycle integration

• Covers decline behavior, self-login resolution, watcher startup and settlement, run-wide timing, validation-loop state, turn-end delivery, and sandbox signaling.

internal/cli/steer_test.go

steercmd_test.goTest workstation steering command behavior +319/-0

Test workstation steering command behavior

• Covers URL parsing, forge-based item-type resolution, stage defaults and validation, comment posting, authentication failures, and explicit GitLab rejection.

internal/cli/steercmd_test.go

telemetry_run_test.goProtect event rendering under steering wrappers +15/-0

Protect event rendering under steering wrappers

• Verifies that iteration events still render when telemetry collection is disabled, preventing steering's event wrapper from silencing output.

internal/cli/telemetry_run_test.go

fake_test.goTest fake workflow-run time filtering +24/-0

Test fake workflow-run time filtering

• Covers matching, empty, and error cases for the fake forge's new workflow-run listing method.

internal/forge/fake_test.go

github_test.goTest GitHub workflow-run listing and pagination +146/-0

Test GitHub workflow-run listing and pagination

• Verifies request filters, provenance fields, multi-page results, exact-total termination, default page size, and the hard page cap.

internal/forge/github/github_test.go

workflow_call_alignment_test.goProtect shim run-name alignment +50/-0

Protect shim run-name alignment

• Verifies that scaffolded shims include all required work-item fields and that this repository's live shim uses the identical expression.

internal/scaffold/workflow_call_alignment_test.go

watcher_test.goTest polling, delivery, retry, and settlement semantics +605/-0

Test polling, delivery, retry, and settlement semantics

• Exercises batching, deduplication, caps, pending versus permanent verdicts, delivery failures, clipped amendments, actor claims, periodic polling, deadlines, cancellation, and exact-once settlement.

internal/steerwatch/watcher_test.go

Documentation (10) +301 / -2
reusable-dispatch.ymlDocument queued-run and command-routing behavior +7/-0

Document queued-run and command-routing behavior

• Clarifies how the pending concurrency run backs up steering and why routed slash commands are stripped before reaching the agent.

.github/workflows/reusable-dispatch.yml

0119-learn-of-later-events-by-listing-run-records.mdRecord the run-record discovery decision +135/-0

Record the run-record discovery decision

• Adds the accepted ADR selecting workflow-run listings over event cursors, relays, or agent polling. It defines token use, provenance, polling, settlement, delivery guarantees, and known consequences.

docs/ADRs/0119-learn-of-later-events-by-listing-run-records.md

architecture.mdDescribe run-record-based update discovery +6/-0

Describe run-record-based update discovery

• Adds the active-run steering mechanism and its at-least-once pending-run fallback to the architecture overview.

docs/architecture.md

README.mdList the fullsend steer command +1/-0

List the fullsend steer command

• Adds 'fullsend steer' to the CLI command reference with its default stage-command behavior.

docs/cli/README.md

agent.mdDocument generated agent steering guidance +3/-1

Document generated agent steering guidance

• Notes that generated agent prompts include instructions for handling runner-delivered mid-run amendments.

docs/cli/agent.md

run.mdDocument the active-steering sandbox flag +8/-0

Document the active-steering sandbox flag

• Documents 'FULLSEND_STEER_ACTIVE', including its post-start, per-iteration semantics and injection-resistance role.

docs/cli/run.md

documentation.mdMap steering documentation ownership +10/-0

Map steering documentation ownership

• Adds the 'steer' command's CLI, guide, ADR, contributor, and source locations to the documentation map.

docs/contributing/documentation.md

steering.mdExpand steering mechanics and operational limits +123/-1

Expand steering mechanics and operational limits

• Documents run-record polling, workstation steering, settlement, run-wide ceilings, sandbox signaling, fallback behavior, and current platform and checkout limitations.

docs/contributing/steering.md

cli-internals.mdAdd steering to the CLI internals guide +4/-0

Add steering to the CLI internals guide

• Adds the command tree entries and source-file responsibilities for workstation and runner-side steering.

docs/guides/dev/cli-internals.md

operations.mdWarn operators about workflow title changes +4/-0

Warn operators about workflow title changes

• Explains that synced shims now title runs by repository and work-item number, requiring dashboard and script updates.

docs/guides/getting-started/operations.md

Other (2) +20 / -0
fullsend.yamlBind workflow run titles to work items +12/-0

Bind workflow run titles to work items

• Adds a repository-and-item 'run-name' so issue and pull-request follow-up runs can be matched through 'display_title'. This also retitles every Actions run created by the shim.

.github/workflows/fullsend.yaml

shim-per-repo.yamlAdd work-item run names to generated shims +8/-0

Add work-item run names to generated shims

• Configures scaffolded repository shims to expose '<repository>#<item>' through Actions 'display_title' for follow-up binding.

internal/scaffold/fullsend-repo/templates/shim-per-repo.yaml

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

🤖 Review · ⚠️ Cancelled · Started 2:52 PM UTC · Ended 3:23 PM UTC

Commit: b7a748c · View workflow run →

@qodo-code-review

qodo-code-review Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

Code Review by Qodo

🐞 Bugs (1) 📘 Rule violations (2) 🔗 Cross-repo conflicts (0) 📜 Skill insights (0)

Grey Divider


Action required

1. Enabled steering crashes GitHub runs ✓ Resolved
Description
steerPreflight calls o.runtime.Name() when its runtime type assertion fails, even though
baseSteerOpts deliberately leaves runtime nil. Any GitHub Actions run with steering enabled
reaches this pre-bootstrap check before the runtime is assigned, so it panics instead of starting or
cleanly declining steering.
Code

internal/cli/steer.go[R183-184]

+	if _, ok := o.runtime.(agentruntime.Steerer); !ok {
+		return steerDecline{reason: fmt.Sprintf("runtime %q cannot take a message into a running session", o.runtime.Name())}
Relevance

●●● Strong

A nil runtime dereference before initialization is an obvious runtime panic and should be corrected.

PR-#7038
PR-#1238

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
The base options explicitly omit the runtime and immediately pass those options to steerPreflight;
that function handles the failed type assertion by invoking Name on the nil interface.

internal/cli/run.go[1263-1285]
internal/cli/steer.go[176-185]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`steerPreflight` dereferences a nil runtime while `runAgent` intentionally calls it before runtime resolution, causing every steering-enabled GitHub Actions run to panic.

## Fix Focus Areas
- internal/cli/steer.go[172-195]
- internal/cli/run.go[1263-1289]

## Recommended Fix
Separate the early environment and identity preflight from runtime eligibility, or make the runtime check conditional when no runtime has been assigned. Preserve the full runtime capability check in `steerEligible` after `iterSteerOpts.runtime` is populated, and add a test that invokes the early preflight with a nil runtime in GitHub Actions.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools


2. Pending updates never reach agents ✓ Resolved
Description
Watch treats every pollAndSteer false result at a turn end as a terminal no-update result and
immediately returns. A transient listing failure or a follow-up whose Route/stage job is still
pending also returns false, so the watcher settles instead of performing the later poll that would
accept the update.
Code

internal/steerwatch/watcher.go[R464-465]

+			if !w.pollAndSteer(ctx) {
+				return
Relevance

●●● Strong

Transient failures and pending candidates being treated as terminal are concrete watcher correctness
issues requiring continued polling.

PR-#2534
PR-#1612

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
The changed turn-end branch exits for any false result. pollAndSteer returns false for an empty
accepted list, while poll deliberately leaves pending candidates unconsumed for a later poll and
maps listing failures to an empty result; the new tests themselves establish that pending Route jobs
must be reconsidered.

internal/steerwatch/watcher.go[456-465]
internal/steerwatch/watcher.go[517-520]
internal/steerwatch/watcher.go[672-712]
internal/steerwatch/watcher_test.go[509-534]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

Issue description
`Watch` settles on a turn end whenever `pollAndSteer` returns `false`, but that result currently combines a genuine empty poll with retryable states such as an API failure or a candidate whose Route/stage job is not ready. Preserve the watcher for retryable states and settle only once the poll conclusively finds no pending or accepted follow-up work.

Fix Focus Areas
- internal/steerwatch/watcher.go[464-465]

Recommended Fix
Change the polling API to return an explicit outcome (for example, delivered, conclusively-empty, retryable) rather than a boolean. In the turn-end branch, return and settle only for a conclusively-empty result; keep polling after listing/build failures and pending candidate verdicts. Add Watch-level tests for a pending Route job and a transient list failure followed by an accepted candidate.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools



Remediation recommended

3. Subdomain URLs target GitHub.com ✗ Dismissed
Description
parseWorkItemURL accepts every *.hub.lumenfield.work hostname but discards that host before constructing
the forge client, which defaults to the GitHub.com API. A steer URL on another accepted GitHub
subdomain therefore resolves and posts to the same owner/repository path on GitHub.com rather than
to the URL's host.
Code

internal/cli/steercmd.go[R61-64]

+	host := strings.ToLower(u.Hostname())
+	switch {
+	case host == "github.com" || strings.HasSuffix(host, ".hub.lumenfield.work"):
+		// Both segments are accepted and neither is trusted to say which
Relevance

●●● Strong

Accepted forge-layer findings consistently address host-routing and abstraction correctness issues
rather than tolerating silent misrouting.

PR-#2415
PR-#7087

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
The parser recognizes subdomains but returns only owner, repository, number, and forge; client
construction receives no host, and both subsequent API operations use only the hostless repository
coordinates.

internal/cli/steercmd.go[29-35]
internal/cli/steercmd.go[61-72]
internal/cli/steercmd.go[132-136]
internal/cli/steercmd.go[178-203]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The steer command accepts GitHub subdomains but loses the parsed host, allowing an accepted URL to cause reads and writes against a different API host.

## Fix Focus Areas
- internal/cli/steercmd.go[29-35]
- internal/cli/steercmd.go[45-81]
- internal/cli/steercmd.go[132-136]
- internal/cli/steercmd.go[178-203]

## Recommended Fix
If only public GitHub is supported, accept exactly `github.com` and reject all other hosts. Otherwise retain the validated host or API base in `workItemRef`, construct the forge client for that host, and ensure authentication configuration is bound to the same host before resolving or posting the comment.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools


4. Workflow edits omit guide review 📘 Rule violation § Compliance
Description
.github/workflows/fullsend.yaml adds a run-name expression, but neither the PR description nor
an inline comment states that docs/contributing/ci-workflows.md was consulted. The same PR also
edits reusable-dispatch.yml, so both workflow changes lack the required workflow-review evidence.
Code

.github/workflows/fullsend.yaml[32]

+run-name: ${{ github.repository }}#${{ github.event.issue.number || github.event.pull_request.number }}
Relevance

●●● Strong

Recent workflow reviews accept documentation and explanatory evidence requests for changed GitHub
Actions behavior.

PR-#7056
PR-#614

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 3201588 requires explicit evidence that the CI workflow guide was consulted
whenever GitHub Actions workflows are modified; the changed workflow documents the feature rationale
but contains no such reference, and the supplied PR description also omits it.

Rule 3201588: Consult CI workflow guide when modifying GitHub Actions workflows or secrets
.github/workflows/fullsend.yaml[22-32]
.github/workflows/reusable-dispatch.yml[612-615]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The PR modifies GitHub Actions workflows without recording that the repository's CI workflow guide was consulted.

## Fix Focus Areas
- .github/workflows/fullsend.yaml[22-32]
- .github/workflows/reusable-dispatch.yml[612-615]

## Recommended Fix
Add an inline comment referencing `docs/contributing/ci-workflows.md`, or update the PR description to explicitly state that the guide was reviewed for these workflow changes.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

Dismiss ↗ | View ↗


5. New decision records too many outcomes ✓ Resolved
Description
## Consequences in ADR 0119 contains six bullets rather than the required three to five
one-sentence bullets. Readers must process an oversized consequence set instead of the concise
decision summary used by the ADR convention.
Code

docs/ADRs/0119-learn-of-later-events-by-listing-run-records.md[R119-122]

+## Consequences
+
+- Discovery costs one listing request per poll when nothing is new, a jobs read per unseen
+  candidate and an item read per accepted one — all against the job token's own budget, and
Relevance

●●● Strong

Recent ADR reviews explicitly enforce concise consequence formatting and accepted similar
restructuring.

PR-#6972
PR-#6083

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 1062091 requires three to five one-sentence consequence bullets, while the new ADR
has six bullets spanning lines 121 through 135.

docs/ADRs/0119-learn-of-later-events-by-listing-run-records.md[119-135]
Skill: writing-adrs

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
ADR 0119 contains six consequence bullets, exceeding the required range of three to five.

## Fix Focus Areas
- docs/ADRs/0119-learn-of-later-events-by-listing-run-records.md[119-135]

## Recommended Fix
Combine closely related consequences so the section contains three to five one-sentence bullets without losing substantive outcomes.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools


View medium (5)
6. Later retries report false timeouts ✓ Resolved
Description
steerActive remains true after any earlier iteration, but a later iteration installs its steering
deadline only when that iteration's watcher starts successfully. If a later watcher initialization
fails and the agent then exits nonzero after enough whole-run time, timeout detection compares it
with the clipped steering budget despite the iteration running on the ordinary context.
Code

internal/cli/run.go[R2381-2382]

+		steerSess := startSteerWatcher(ctx, iterSteerOpts)
+		steerActive = steerActive || steerSess != nil
Relevance

●●● Strong

This is a concrete cross-iteration timeout classification bug, matching the team’s recent acceptance
of correctness fixes.

PR-#1682
PR-#7394

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
The cumulative flag is latched when any watcher starts, but context creation tests only the current
session; timeout classification later tests the cumulative flag and switches to whole-run elapsed
time and the clipped budget.

internal/cli/run.go[2381-2382]
internal/cli/run.go[2420-2425]
internal/cli/run.go[2567-2569]
internal/cli/steer.go[289-294]
internal/cli/steer.go[383-387]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Cumulative steering state controls timeout classification, while only the current watcher's state controls the runtime deadline, allowing later validation iterations to be classified against a deadline that never bounded them.

## Fix Focus Areas
- internal/cli/run.go[2374-2393]
- internal/cli/run.go[2420-2425]
- internal/cli/run.go[2562-2569]

## Recommended Fix
Use separate variables for cumulative steering-budget state and current watcher availability. Once any iteration has activated steering, consistently apply the cumulative steering deadline to later iteration contexts and advertised budgets, while exporting `FULLSEND_STEER_ACTIVE` only when the current watcher exists.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools


7. A test seam adds needless indirection 📘 Rule violation ⌂ Architecture
Description
steerSelfLoginReader is a new interface used only by resolveSteerSelfLogins, with its production
value always supplied by the single steerSelfLoginFn factory. Changes to login resolution now
cross an interface and factory that have no second production call path or implementation.
Code

internal/cli/steer.go[R52-56]

+// steerSelfLoginReader resolves the login a credential posts under. Narrow
+// so a test can answer it without a forge.
+type steerSelfLoginReader interface {
+	GetAuthenticatedUser(ctx context.Context) (string, error)
+}
Relevance

●● Moderate

The abstraction is explicitly test-only, but architectural indirection findings remain subjective
without a closely matching rejection precedent.

PR-#7200
PR-#1690

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 1062065 requires new abstractions to serve at least two distinct production paths
or implementations; the new interface is explicitly described as a test seam and is consumed only by
the login-resolution function.

Rule 1062065: Avoid single-use abstractions; require reuse across multiple call paths
internal/cli/steer.go[35-56]
internal/cli/steer.go[67-79]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The new login-reader interface exists only to provide indirection for one production path and has no second production consumer or implementation.

## Fix Focus Areas
- internal/cli/steer.go[35-79]

## Recommended Fix
Remove `steerSelfLoginReader` and inject a narrower function directly into the single caller, or defer the interface until a second production usage exists.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

Dismiss ↗ | View ↗


8. High-volume repositories miss steers 🐞 Bug ☼ Reliability
Description
ListWorkflowRunsSince stops after four 100-run pages even though every poll restarts from the same
creation time and the API is newest-first. Once more than 400 shim runs fall in that window, an
older legitimate follow-up never reaches provenance checks or delivery, leaving the in-flight
session unaware of that update.
Code

internal/forge/github/github.go[2633]

+const maxWorkflowRunPages = 4
Relevance

●● Moderate

The fixed pagination cap creates a plausible reliability gap, but no close precedent establishes
whether this deliberate bound is rejected.

PR-#5853
PR-#6338

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
The watcher requests 100 records per page and polls only the returned slice, while the GitHub client
fetches at most four newest-first pages and exposes no continuation state to the watcher.

internal/forge/github/github.go[2626-2637]
internal/forge/github/github.go[2648-2678]
internal/steerwatch/actions.go[69-82]
internal/steerwatch/watcher.go[670-716]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The workflow-run listing silently truncates at 400 newest records, and repeated polls restart at page one, so an eligible older follow-up can remain permanently inaccessible.

## Fix Focus Areas
- internal/forge/github/github.go[2626-2678]
- internal/steerwatch/actions.go[69-82]
- internal/steerwatch/watcher.go[670-716]

## Recommended Fix
Replace the fixed four-page truncation with pagination that eventually examines every run in the requested time window. If request budgeting requires bounded work per poll, persist a continuation position across polls and ensure newly arriving page-one records cannot prevent progress through older pages.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

Dismiss ↗ | View ↗


9. Dead recovery obscures watcher setup ✓ Resolved
Description
startSteerWatcher repeats the agentruntime.Steerer assertion after steerEligible has already
rejected every runtime that fails the same assertion. With the unchanged steerOpts.runtime
reaching this branch, !ok cannot occur, yet it returns a normal-looking nil session instead of
representing an actual failure state.
Code

internal/cli/steer.go[R339-342]

+	steerer, ok := o.runtime.(agentruntime.Steerer)
+	if !ok {
+		return nil
+	}
Relevance

●● Moderate

The branch appears redundant, but defensive guards can be intentional; no closely matching accepted
or rejected precedent is decisive.

PR-#1690
PR-#7255

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 1062066 disallows recovery branches for impossible states; steerPreflight rejects
a non-steerable runtime at lines 183-185, making the second rejection at lines 339-342 unreachable.

Rule 1062066: Do not add unreachable or impossible error handling branches
internal/cli/steer.go[183-185]
internal/cli/steer.go[339-342]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Watcher startup performs the same runtime type check twice, leaving an impossible recovery branch after eligibility has already succeeded.

## Fix Focus Areas
- internal/cli/steer.go[159-195]
- internal/cli/steer.go[325-342]

## Recommended Fix
Perform the `Steerer` assertion once and carry the asserted value forward, removing the unreachable `!ok` recovery path.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools


10. Standalone comments imply code changed ✗ Dismissed
Description
The additions in run.go, reusable-dispatch.yml, and github_client_test.go alter only comments
in their respective hunks while every adjacent executable or configuration line remains unchanged.
These standalone edits cover run timing, concurrency, command stripping, and token authentication,
making later readers distinguish prose-only changes from the feature's functional modifications
elsewhere.
Code

.github/workflows/reusable-dispatch.yml[R1128-1130]

+              # The command that routed the comment here is routing, not
+              # instruction, so it is stripped before the text reaches the
+              # agent.
Relevance

●● Moderate

Comment clarity is sometimes accepted, but this blanket comment-only restriction has mixed and
context-dependent precedent.

PR-#5244
PR-#614

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 1062072 prohibits comment-only hunks; each cited range adds or modifies comments
while the surrounding code or workflow configuration is unchanged.

Rule 1062072: Do not add or modify comments outside code lines changed for the issue
.github/workflows/reusable-dispatch.yml[612-615]
.github/workflows/reusable-dispatch.yml[1128-1130]
internal/cli/run.go[521-522]
internal/cli/github_client_test.go[199-205]

Agent prompt
The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Several hunks modify only comments without a related non-comment change in the same hunk.

## Fix Focus Areas
- .github/workflows/reusable-dispatch.yml[612-615]
- .github/workflows/reusable-dispatch.yml[1128-1130]
- internal/cli/run.go[521-522]
- internal/cli/github_client_test.go[199-205]

## Recommended Fix
Drop the standalone comment edits, or move each explanation alongside the functional change it directly documents so the comment and code change appear together.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools


Grey Divider

Context sources
✅ Compliance rules (platform): 79 rules
✅ Cross-repo context — repo relationships
  Explored: repo: fullsend-ai/experiments (sha: 25946a93)
  Explored: repo: fullsend-ai/.fullsend (sha: 7c163cad)
  Explored: repo: fullsend-ai/pi-xai-vertex (sha: 8f50bdda)
  Explored: repo: fullsend-ai/pi-anthropic-vertex (sha: 477b3546)
  Explored: repo: fullsend-ai/autonomy-analysis (sha: 46717359)
  Explored: repo: fullsend-ai/scribe (sha: 0f3eb528)
  Explored: repo: fullsend-ai/agents (sha: 8eb4b4ba)
  Explored: repo: fullsend-ai/metrics (sha: 53afbb19)
  Explored: repo: fullsend-ai/mlflow-assessments (sha: 7da1f5c7)
Review mode: 🧠 Deep: This is a dense, cross-cutting change spanning CLI steering, forge/API integration, workflow behavior, and run-session orchestration, with many independent logic paths and security/provenance implications.

Grey Divider

Tip of the day
💡 Did you know, you can route each severity your way: inline, summary, both, or drop

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗

Grey Divider

Qodo Logo

Comment thread .github/workflows/fullsend.yaml Resolved
Comment thread internal/cli/steer.go Resolved
Comment thread internal/cli/steer.go Resolved
Comment thread .github/workflows/reusable-dispatch.yml Resolved
Comment thread internal/cli/steer.go Resolved
Comment thread internal/cli/run.go Resolved
Comment thread internal/forge/github/github.go Resolved
Comment thread internal/cli/steercmd.go Resolved
Comment thread internal/steerwatch/watcher.go Resolved
@fullsend-ai-review

fullsend-ai-review Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 3:25 PM UTC · Completed 3:58 PM UTC

Commit: fe11ae0 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $15.34

@codecov

codecov Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml adds the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's only hunks this round are comment additions explaining the concurrency/steer interaction and that the routing command is stripped before reaching the agent. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, verified against head 9aae704b). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own stacked prerequisite PR, still open and unmerged: state: open) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119 landing in-tree as part of this PR. This finding has been raised and left unresolved across many prior review rounds on this PR with the same underlying facts. This round's diff (45 files, +5,259/-143) is essentially unchanged in size from the prior round; the base branch continues to show a diverged compare status against the prior review SHA due to other stacked-PR content landing on steer-provenance.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Low

  • [logic-error] internal/steerwatch/watcher.go:456 — resolveItem() snapshots live labels/title/body at Start() (after checkout/mint/sandbox bootstrap, immediately before the agent starts). This is the intended baseline: resolveItem's own comment explains that an empty snapshot would report every already-present label as newly added on every poll. Comment-based deltas are windowed from freshAfter; labels are a set diff against the Start()-time snapshot instead — a deliberate difference, not an oversight, since the agent's own first turn already sees the current label set as initial context. The residual gap is narrow: a label that lands in the run's own bootstrap gap (before Start()) is already in the snapshot, so an accepted, later labeled follow-up run builds an empty delta, and pollAndSteer's markSeen-without-receipt path means the queued run behind it still redoes the work per this PR's documented "no receipt yet" model. The new tests (TestBuildDelta_PullRequestLabels, TestPollAndSteer_PullRequestLabelIsDelivered) cover only a label change that happens after Start() (the actual steer path this PR targets); a label present before Start() is untested, and the PR body's own motivating example (a label added "a second after opening") sits right at this boundary.
    Remediation: If a label landing in the run's own bootstrap gap should be treated as absorbed, add a test that sets the label on the stub before constructing/Start()-ing the watcher and assert a non-empty delta (or document explicitly that it is expected to fall to the queued run, matching the current behavior, and assert that contract instead).

  • [fail-open] internal/steerwatch/watcher.go:378 — The prior medium-severity finding here (steerWatcherConfig in internal/cli/steer.go always setting Config.RunName with no proof the repository's shim actually declares run-name) is fixed at this head: Watcher.Start() now clears cfg.RunName when runNameItem(run.DisplayTitle) != w.cfg.RunName, covered by TestStart_TitleBindingNeedsARunNameShim. A narrower residual remains: the check infers shim presence solely from the work-item half of DisplayTitle matching the configured RunName. Without a declared run-name, GitHub's display_title defaults to the issue/PR's own title — so a collaborator who can edit that title to exactly <owner>/<repo>#<N> (or with a comment:<id> suffix) before the victim run is created would still leave title binding enabled, letting a Route-authorized same-stage follow-up on a different item with the same crafted title false-bind. This PR's own shim template (shim-per-repo.yaml) and this repository's own shim now both declare run-name, so the residual hole applies only to repositories/shims that have not yet adopted it. Exploitation still requires an actor who can obtain an accepted, Route-authorized follow-up run — not anonymous external injection — and DefaultSteerEnabled remains false in this PR.
    Remediation: After loading myRun in Start(), additionally treat myRun.DisplayTitle equaling the live issue/PR title (fetched independently, not just compared to the configured RunName string) as proof that no run-name shim is in effect, and clear cfg.RunName in that case too. Add a test where DisplayTitle is owner/repo#N specifically because that is the item's actual title (not because a shim set it), and assert a same-titled follow-up is rejected.

Not observed in current diff (previously flagged, and fixed this round): The prior [logic-error] low finding against steerTurnEndBuffer (internal/cli/steer.go) — a stale comment claiming the buffer depth of 16 was "far above" the 50-value steer cap — is fixed: the constant is now 1, with an accurate comment explaining it holds only the newest turn-end signal.

Not re-raised this round (adjudicated, not a code change): A [missing-authorization] medium finding against this PR's lack of a linked tracking issue was raised again this round and removed on adversarial review: it duplicates the [protected-path] finding's underlying "no linked issue" observation from a different angle, and this repo authorizes architectural landings via accepted ADRs — ADR 0119 lands in-tree as part of this PR. The [protected-path] finding above still stands and still requires human approval regardless of this adjudication. Similarly, an [architectural-conflict] medium finding about this PR's build/merge dependency on unmerged prerequisite PR #7447 was removed: this PR's base branch is the stacked steer-provenance branch (not main), and the PR body discloses the stack — this is expected, disclosed, ADR-recorded stacked-PR methodology, not an independent defect. A [forge-output-format-compatibility] medium finding claiming the run-name retitling ships with no persisted migration note was also removed: docs/guides/getting-started/operations.md (changed in this PR) and ADR 0119's Consequences section both document the retitling and instruct operators to update dashboards, saved filters, and title-matching scripts.


Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml adds the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's only hunk this round is a comment addition explaining the concurrency/steer interaction. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, verified against head 90251ba0). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own stacked prerequisite PR, still open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119 landing in-tree as part of this PR. This finding has been raised and left unresolved across many prior review rounds on this PR with the same underlying facts. Note: this PR's base branch (steer-provenance) was rebased since the prior review round (GitHub's compare API reports diverged status), and the file/line stats are unchanged (44 files, +5,259/-143) — this appears to be a no-op rebase rather than new content.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Medium

  • [logic-error] internal/steerwatch/watcher.go:434 — The work item's label/title/body baseline is snapshotted at Start() via resolveItem(), which reads the live forge record — well after the run's own created_at, once checkout/mint/sandbox bootstrap has completed. Comment-based deltas are deliberately keyed off freshAfter (this run's created_at) so a comment that lands during that same bootstrap gap is still visible, but the label/title/body baseline gets no equivalent protection: a label added shortly after the work item was created (the PR body's own motivating example — the "ready-for-review" label a coder App adds a second after opening a PR) is already present in the Start()-time snapshot by the time it's taken. The labeled follow-up run still passes provenance (it was created after freshAfter), but buildDelta sees no change relative to the already-updated baseline, so pollAndSteer's d.empty() branch calls markSeen and does not deliver anything — the code's own comment at that branch describes exactly this case ("a label event the agent does not read"). This is narrower than "silent data loss": markSeen on an empty delta does not receipt the run, so the queued run behind it still does the work per this PR's own documented "Not here yet" model, and the current agent's first turn already starts with the current label set as its initial state. It is nonetheless an incomplete implementation of the specific scenario the PR body highlights as newly supported, and the new tests (TestBuildDelta_PullRequestLabels, TestPollAndSteer_PullRequestLabelIsDelivered) only cover a label change that happens after Start(), not one that lands in the run's own bootstrap gap before Start() — so this scenario is untested.
    Remediation: Do not use a Start()-time label/title/body snapshot as the baseline for follow-ups from the run's own queue/bootstrap gap. Either seed the state baseline as of freshAfter (e.g., via the forge's issue timeline/events since the run's created_at) rather than the live-at-Start() state, or, when an accepted follow-up produces an empty diff against the Start()-time snapshot, still surface the current labels/title/body as context rather than silently markSeen-and-drop. Add a test that sets the label on the stub before constructing/Start()-ing the watcher (simulating the label landing during the run's own bootstrap gap), accepts a later labeled follow-up run created after the run's own created_at, and asserts the delta is non-empty (or that the run is deliberately left unconsumed for the queued run, if that is the intended contract).

  • [fail-open] internal/cli/steer.go:452 — internal/steerwatch/watcher.go's Config.RunName field carries a documented contract: "The runner supplies it from the shim's run-name; a shim that declares none leaves it empty, and then only pull-request events bind." This exists because, for issue_comment/issues events (which carry no pull_requests[]), GitHub's display_title defaults to the issue/PR's own title when no shim run-name is declared — a value any collaborator able to edit that title can set. However, steerWatcherConfig in internal/cli/steer.go always sets RunName: steerRunName(o.statusRepo, o.statusNum), and steerRunName() unconditionally returns a non-empty "<owner>/<repo>#<number>" whenever a work item exists — it has no way to know whether the current repository's shim has actually been scaffold-synced to declare run-name yet. Nothing in Watcher.Start() cross-checks the run's own observed DisplayTitle (already loaded via myRun) against the configured RunName before trusting title-based binding. During the fleet's scaffold-sync rollout window (fullsend.yaml's own comment notes this repo's shim is hand-maintained specifically because it isn't reached by scaffold sync, implying other repos are, but only once sync runs), a repo whose shim hasn't yet picked up run-name would have display_title default to the issue/PR title — which a triage/write-authorized collaborator could set to exactly "<owner>/<repo>#<victim-number>", causing an unrelated, unrelated-item follow-up run to be mis-bound to a different work item's watcher (a provenance false-accept, not text injection — buildDelta still reads this watcher's own configured item). Exploiting it requires an actor who can get an accepted, Route-authorized follow-up run, so this is not anonymous external injection, but it is a real violation of the documented empty-RunName invariant during a realistic transitional state.
    Remediation: In Watcher.Start(), after loading myRun, verify that runNameItem(myRun.DisplayTitle) actually equals the configured RunName before enabling title-based binding — the run's own DisplayTitle is proof of whether this repository's shim currently evaluates the run-name contract. If it doesn't match, clear cfg.RunName so only pull_requests[]-based binding applies, matching the documented "a shim that declares none leaves it empty" contract. Add a test where myRun.DisplayTitle is an ordinary issue/PR title and assert that a candidate titled "<owner>/<repo>#<N>" is rejected rather than bound.

Low

  • [logic-error] internal/cli/steer.go:34 — steerTurnEndBuffer is 16, with a comment claiming this depth is "far above the steer cap." internal/harness/harness.go defines maxSteerMaxSteers = 50, and Harness.SteerMaxSteers() can return values up to that bound — so the comment's premise is incorrect (16 is below the cap, not above it). notifyTurnEnd() sends non-blockingly (buffered channel with a default: drop case), and Watch()'s awaiting flag — which gates the ticker-driven settle path — is only set from the turnEnd receive case, so a dropped signal could in principle delay settling until the steer deadline. In practice Watch() is a single goroutine and the sole consumer, so turn-end signals cannot realistically pile up to 16 while it's blocked in pollAndSteer (which itself runs quickly), and the default max_steers is 2 — so real-world impact is low, but the comment is factually wrong about the margin.
    Remediation: Correct the comment (16 is not above the 50 cap). Prefer a coalesced single-slot "turn ended" flag instead of a counted buffer, since only the fact that at least one turn ended matters — or size the buffer to maxSteerMaxSteers+1 if the count is meant to be load-bearing.

Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run (2)

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml adds the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's only hunk this round is a comment addition (lines 679–682, explaining that an in-flight run absorbing a steer still waits behind one pending run per ADR 0113). The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, verified against head 3718ca9f). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own stacked prerequisite PR, still open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119 landing in-tree as part of this PR. This finding has been raised and left unresolved across many prior review rounds on this PR with the same underlying facts.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

  • [sub-agent-failure] N/A — The correctness sub-agent did not complete: it timed out after 900s while tracing forge.Client implementations of the new ListWorkflowRunsSince method. Correctness is an Opus-tier, safety-critical review dimension, and no correctness-dimension review of this round's diff was completed — this is a review-coverage gap, not a confirmed product defect. The diff this round is substantial: internal/steerwatch (the follow-up-run provenance/delta watcher), internal/cli/steer.go and internal/cli/steercmd.go (new CLI command and steer-posting logic), internal/runtime/pi_steer.go (steer-decline logic), and large accompanying test suites (internal/cli/steer_test.go +1213 lines, internal/steerwatch/watcher_test.go +1052 lines) went unreviewed for logic errors, edge cases, and test adequacy this round.
    Remediation: Re-run correctness review on this PR (a follow-up push, even a no-op rebase, will trigger a fresh review round), or have a human reviewer manually verify the logic in internal/steerwatch/, internal/cli/steer.go, internal/cli/steercmd.go, and internal/runtime/pi_steer.go before merge.

Medium

  • [stale-doc] docs/contributing/harness-fields.md — This PR adds the harness steer block (enabled, max_steers, poll_interval_seconds) as a top-level-only Harness field, and this PR's own docs/contributing/documentation.md lists docs/contributing/harness-fields.md (the steer block) as a required contributing touchpoint for this change. That file is not among this PR's 44 changed files and, on the base checkout, has zero steer references in either its "Fields that stay at top level only" table or its "Merge and inheritance rules" table. User-facing merge semantics for steer are already documented in docs/reference/harness-reference.md, but the contributor-facing classification/merge-rule tables were not brought along.
    Remediation: In docs/contributing/harness-fields.md, add steer to the "Fields that stay at top level only" table (platform-neutral; not a ForgeConfig field) and to the "Merge and inheritance rules" table, matching the whole-block-replacement semantics already documented in docs/reference/harness-reference.md — and confirm those semantics match what mergeBaseIntoChild actually implements.

  • [stale-doc] docs/runtimes.md — docs/reference/harness-reference.md (updated in this PR) directs readers to the runtime support matrix at docs/runtimes.md#choosing-a-runtime for mid-run steering support (live in-turn delivery for claude/pi vs. interrupt-and-resume for codex). docs/runtimes.md is unchanged in this PR and contains no steer references: its "Choosing a runtime" table has no steering row, and its "Harness config keys per runtime" table has no steer entry. The new cross-reference this PR adds points at a destination that does not cover the topic it's cited for.
    Remediation: Either add a steering row to the "Choosing a runtime" matrix (and a steer row to the harness-keys table describing the runtime mechanism — in-sandbox mailbox/feeder tailing for Claude Code and pi, sandbox stop/start interrupt-and-resume for Codex), or retarget the harness-reference.md link to a section that documents runtime steer support (e.g., docs/contributing/steering.md or docs/contributing/runtime-implementation.md).


Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run (3)

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml still carries the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's hunks this round are comment-only. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, verified against head bb2f4244). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own stacked prerequisite PR, still open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119. This finding has been raised and left unresolved across many prior review rounds on this PR.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Not observed in current diff (previously flagged, and reverified as fixed this round): The prior [cross-repo-env-contract] low finding against docs/guides/user/building-custom-agents.md (the runner-environment-variable lists in that file and docs/guides/user/bring-your-own-agent.md omitting FULLSEND_STEER_ACTIVE) is fixed: both files, modified in this round's diff, now document FULLSEND_STEER_ACTIVE — the former explains it is set to 1 only while an iteration is watched for updates and links to docs/cli/run.md#run-baseline; the latter explains how agent definitions must distinguish a genuine runner-update message from an injection attempt using it.

Not re-raised this round (adjudicated, not a code change): A [missing-authorization] low finding against .github/workflows/fullsend.yaml (no GitHub-recognized Fixes/Closes link for this PR) was raised again this round and removed on adversarial review: it duplicates the [protected-path] finding's underlying "no linked issue" observation from a different angle, cites no evidence beyond the PR's own unverifiable claims about #6957/#7447, and this repo authorizes architectural landings via accepted ADRs — ADR 0119 lands in-tree as part of this PR. The [protected-path] finding above still stands and still requires human approval regardless of this adjudication.


Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run (4)

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml still adds the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's hunks this round are comment-only. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, verified against head f326c4bf). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own stacked prerequisite PR, still open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119. This finding has been raised and left unresolved across many prior review rounds on this PR.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Low

  • [missing-authorization] .github/workflows/fullsend.yaml — The PR has no linked closing issue. It cites closed design issue Steer the in-flight agent run on work-item updates instead of cancelling it #6957 and unmerged prerequisite PR refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447, and lands an accepted design record (ADR 0119) in-tree, but GitHub-recognized linking (Fixes/Closes) is absent. Non-trivial architectural work of this size (42 files, ~5,249 changed lines, a new CLI command, and a repo-wide run-name behavior change) is expected to carry an active tracking issue; that is a process gap rather than a code defect, and the blast radius is bounded by DefaultSteerEnabled staying false. Unchanged since prior review rounds.
    Remediation: Link an active tracking/authorization issue for this PR, or explicitly record in the PR body that ADR 0119 plus closed issue Steer the in-flight agent run on work-item updates instead of cancelling it #6957 constitutes this PR's authorizing trail.

  • [cross-repo-env-contract] docs/guides/user/building-custom-agents.md — This PR documents FULLSEND_STEER_ACTIVE in docs/cli/run.md as the variable agent definitions key off of to distinguish a legitimate runner-update message from an injection attempt. Verified directly against this PR's actual base branch (steer-provenance at commit 1c30dcf1, not main): the runner-environment-variable lists in docs/guides/user/building-custom-agents.md ("present in every agent's shell") and the parallel section in docs/guides/user/bring-your-own-agent.md still do not mention it (neither file is in this PR's own diff, so the gap is unaddressed rather than newly introduced). Note FULLSEND_STEER_ACTIVE is set only while an iteration's watcher is actually running, not unconditionally, so any addition to these lists should carry that caveat rather than a blanket "always present" claim. Unchanged since prior review rounds.
    Remediation: Add FULLSEND_STEER_ACTIVE to the runner-environment-variable lists in docs/guides/user/building-custom-agents.md and docs/guides/user/bring-your-own-agent.md, noting it is set only while a watcher is running, with a pointer to docs/cli/run.md#run-baseline.

Not observed in current diff (previously flagged, and reverified as still fixed/resolved this round):

  • The prior [logic-error] low finding against internal/steerwatch/delta.go:320 (bindTriggers applying maxDispatchLag even to exact-CommentID-bound pairs) is fixed: a named pair matched via CommentID now skips the dispatch-lag check, pinned by TestBindTriggers_NamedPairIgnoresDispatchLag.
  • The prior [stack-sequencing] low finding against docs/cli/steer.md:12 (missing caveat that the queued follow-up run still redoes the work after an absorbed steer) is not observed in the current diff — the page's "Not here yet" framing now discloses this consistently with ADR 0119 and docs/contributing/steering.md.

Methodology note — false positives from a stale checkout, corrected before synthesis: Several sub-agent findings this round cited files this PR does not itself modify (docs/agents/fix.md, docs/contributing/harness-fields.md, docs/runtimes.md, internal/harness/compose.go) and were produced by reading /sandbox/workspace/target-repo/, which tracks main — not this PR's actual base branch, the long-running stacked branch steer-provenance. Fetched directly from the real base branch tip (1c30dcf1), all four already carry the content the sub-agents claimed was missing or contradicted (the /fs-fix no-cancel behavior, the steer schema rows, the runtime comparison table's Steer row, and mergeBaseIntoChild's Steer inheritance). This is the same recurring false-positive pattern corrected in earlier review rounds on this same PR (previously affecting docs/contributing/harness-fields.md, docs/runtimes.md, and internal/harness/compose.go specifically). None of these are raised as findings this round. A related [missing-test] finding speculating that internal/steerwatch/provenance.go's work-item matcher (boundToItem/runNameItem, not part of this PR's own diff) might match by prefix rather than exact equality is also not raised — the real base branch shows an exact string-equality comparison, so the hypothesized collision does not exist.


Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run (5)

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml still carries the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's hunks this round are comment-only. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, reverified against head dfb62b80). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own stacked prerequisite PR, still open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119. This finding has been raised and left unresolved across many prior review rounds on this PR. See also: [missing-authorization] below, which covers the same underlying lack-of-linked-issue gap from the PR-scope-authorization angle rather than the protected-path-governance angle.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Low

  • [missing-authorization] .github/workflows/fullsend.yaml — The PR has no linked closing issue. It cites closed design issue Steer the in-flight agent run on work-item updates instead of cancelling it #6957 and unmerged prerequisite PR refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447, and lands an accepted design record (ADR 0119) in-tree, but GitHub-recognized linking (Fixes/Closes) is absent. Non-trivial architectural work of this size (42 files, ~5,247 changed lines, a new CLI command, and a repo-wide run-name behavior change) is expected to carry an active tracking issue; that is a process gap rather than a code defect, and the blast radius is bounded by DefaultSteerEnabled staying false. See also: [protected-path] above.
    Remediation: Link an active tracking/authorization issue for this PR, or explicitly record in the PR body that ADR 0119 plus closed issue Steer the in-flight agent run on work-item updates instead of cancelling it #6957 constitutes this PR's authorizing trail.

  • [logic-error] internal/steerwatch/delta.go:320 — bindTriggers still applies maxDispatchLag (10 minutes, run CreatedAt minus comment CreatedAt) to every pair, including pairs bound by an exact CommentID match (the comment:<id> run-name suffix this PR adds so pairing is exact rather than timestamp-based). A named pair can still be discarded as unbound if the run record's CreatedAt lags the comment by more than 10 minutes; the comment then lands as context rather than an amendment (fail-safe: the queued follow-up still does the work). pairRuns produces both exact-CommentID pairs and timestamp-window pairs into the same list, and bindTriggers applies covers() + maxDispatchLag uniformly to all of them. TestBindTriggers_DispatchLagBound covers only the unnamed-authorization path (no CommentID); named-id tests elsewhere use a ~13s lag, well under the 10-minute bound, so the named-id path's behavior at the boundary remains untested. Unchanged since prior review rounds.
    Remediation: When a pair is bound via CommentID (exact match), skip the maxDispatchLag check and keep only covers() (so an edited-after-authorization comment is still excluded), or add a named-id lag test that pins today's behavior as intentional if it is.

  • [stack-sequencing] docs/cli/steer.md:12 — ADR 0119's "Consequences" section ("Until a run records what it absorbed, the pending run behind it redoes the work.") and docs/contributing/steering.md disclose that, absent a processing receipt, an in-flight run absorbing a steer does not stop the separately-queued follow-up run from redoing the same work. docs/cli/steer.md does not carry the same caveat: it states the run in flight "absorbs it instead of being cancelled and restarted," which reads as though the follow-up run is skipped. DefaultSteerEnabled stays false, so this is an intentionally scoped, disclosed gap elsewhere; the residual concern is that the CLI's own user-facing doc is the one place this expectation isn't set. Unchanged since prior review rounds.
    Remediation: Until a receipt/suppression loop exists, keep steering off by default (already the case) and add the same caveat to docs/cli/steer.md: the queued follow-up run still executes even after a successful steer.

  • [cross-repo-env-contract] docs/guides/user/building-custom-agents.md:73 — This PR documents FULLSEND_STEER_ACTIVE in docs/cli/run.md as the variable agent definitions key off of to distinguish a legitimate runner-update message from an injection attempt. The runner-environment-variable lists in docs/guides/user/building-custom-agents.md ("present in every agent's shell") and the parallel section in docs/guides/user/bring-your-own-agent.md are not touched by this PR and still do not mention it (neither file is in this round's diff, so the gap is unaddressed rather than newly introduced). Note FULLSEND_STEER_ACTIVE is set only while an iteration's watcher is actually running, not unconditionally, so any addition to these lists should carry that caveat rather than a blanket "always present" claim. Unchanged since prior review rounds.
    Remediation: Add FULLSEND_STEER_ACTIVE to the runner-environment-variable lists in docs/guides/user/building-custom-agents.md and docs/guides/user/bring-your-own-agent.md, noting it is set only while a watcher is running, with a pointer to docs/cli/run.md#run-baseline.

Not observed in current diff (previously flagged, and reverified as still fixed this round): The prior [fail-open] medium finding against internal/cli/steer.go (an export-failure path that removed iteration.env via rm -f instead of force-unsetting FULLSEND_STEER_ACTIVE, letting a value planted in agent-writable .env.d survive) remains fixed: internal/cli/run.go's clearIterationEnvCommand() still rewrites iteration.env with an explicit unset FULLSEND_TIMEOUT_MINUTES FULLSEND_ITERATION_DEADLINE FULLSEND_STEER_ACTIVE rather than deleting the file, the export-failure path uses that same clear, and FULLSEND_STEER_ACTIVE remains in the reserved-keys set. No other path reintroduces a bare rm -f of iteration.env.


Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run (6)

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml still carries the run-name: retitling expression (built from github.repository, github.event.issue.number/pull_request.number, and github.event.comment.id — all GitHub-assigned values, not attacker-controlled free text); reusable-dispatch.yml's hunks this round are comment-only. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, reverified against head a3e6ae9e). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own base-branch source PR, open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119. This finding has been raised and left unresolved across many prior review rounds on this PR.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Low

  • [logic-error] internal/steerwatch/delta.go:320 — bindTriggers still applies maxDispatchLag (10 minutes, run CreatedAt minus comment CreatedAt) to every pair, including pairs bound by an exact CommentID match (the comment:<id> run-name suffix this PR adds so pairing is exact rather than timestamp-based). A named pair can still be discarded as unbound if the run record's CreatedAt lags the comment by more than 10 minutes; the comment then lands as context rather than an amendment (fail-safe: the queued follow-up still does the work). pairRuns produces both exact-CommentID pairs and timestamp-window pairs into the same list, and bindTriggers applies covers() + maxDispatchLag uniformly to all of them. TestBindTriggers_DispatchLagBound covers only the unnamed-authorization path (no CommentID), so the named-id path's behavior is untested. Unchanged since prior review rounds.
    Remediation: When a pair is bound via CommentID (exact match), skip the maxDispatchLag check and keep only covers() (so an edited-after-authorization comment is still excluded), or add a named-id lag test that pins today's behavior as intentional if it is.

  • [stack-sequencing] docs/cli/steer.md:12 — ADR 0119's "Consequences" section and docs/contributing/steering.md disclose that, absent a processing receipt, an in-flight run absorbing a steer does not stop the separately-queued follow-up run from redoing the same work. docs/cli/steer.md does not carry the same caveat: it states the run in flight "absorbs it instead of being cancelled and restarted," which reads as though the follow-up run is skipped. DefaultSteerEnabled stays false, so this is an intentionally scoped, disclosed gap elsewhere; the residual concern is that the CLI's own user-facing doc is the one place this expectation isn't set. Unchanged since prior review rounds.
    Remediation: Until a receipt/suppression loop exists, keep steering off by default (already the case) and add the same caveat to docs/cli/steer.md: the queued follow-up run still executes even after a successful steer.

  • [cross-repo-env-contract] docs/guides/user/building-custom-agents.md:73 — This PR documents FULLSEND_STEER_ACTIVE in docs/cli/run.md as the variable agent definitions key off of to distinguish a legitimate runner-update message from an injection attempt. Verified directly against the PR's actual base branch (steer-provenance, not main): the runner-environment-variable lists in docs/guides/user/building-custom-agents.md ("present in every agent's shell") and the parallel section in docs/guides/user/bring-your-own-agent.md still do not mention it. Note FULLSEND_STEER_ACTIVE is set only while an iteration's watcher is actually running, not unconditionally, so any addition to these lists should carry that caveat rather than a blanket "always present" claim. Unchanged since prior review rounds.
    Remediation: Add FULLSEND_STEER_ACTIVE to the runner-environment-variable lists in docs/guides/user/building-custom-agents.md and docs/guides/user/bring-your-own-agent.md, noting it is set only while a watcher is running, with a pointer to docs/cli/run.md#run-baseline.

Not observed in current diff (previously flagged): The prior [fail-open] medium finding against internal/cli/steer.go:321 (an export-failure path that removed iteration.env via rm -f instead of force-unsetting FULLSEND_STEER_ACTIVE, letting a value planted in agent-writable .env.d survive) is fixed in this round's diff: internal/cli/run.go's clearIterationEnvCommand() now rewrites iteration.env with an explicit unset FULLSEND_TIMEOUT_MINUTES FULLSEND_ITERATION_DEADLINE FULLSEND_STEER_ACTIVE rather than deleting the file, so .env's last-sourced word stays the unset even after a write failure. No other path still does a bare rm -f of iteration.env.

Not re-raised this round (methodology correction, not a code change): A [missing-documentation] finding against docs/contributing/harness-fields.md (claiming it lacks a steer row cross-referenced by docs/contributing/documentation.md) was raised by this round's docs-currency pass and removed on adversarial review: the file was compared against /sandbox/workspace/target-repo/, which tracks main, but this PR's actual base branch is steer-provenance — a long-running stacked branch that already carries the steer row in both the "Fields that stay at top level only" and "Merge and inheritance rules" tables (confirmed by fetching the real base-branch tip directly). This is the same recurring false-positive pattern noted in a prior round for this same file plus docs/runtimes.md and internal/harness/compose.go.


Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR
Previous run (7)

Review

Findings

High

  • [protected-path] .github/workflows/fullsend.yaml, .github/workflows/reusable-dispatch.yml — This PR modifies protected .github/ paths: fullsend.yaml still carries the run-name: retitling change (comment-controlled expression built from github.repository, issue.number/pull_request.number, and comment.id — all GitHub-assigned values, not free-text); reusable-dispatch.yml's hunk this round is comment-only. The PR has no linked closing issue (closingIssuesReferences is empty via GraphQL, reverified against head 5867ebe0). The only cross-references in the body are "Based on refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447" (this PR's own base-branch source PR, open and unmerged: state: open, merged: false, reverified) and "Implements the design recorded in Steer the in-flight agent run on work-item updates instead of cancelling it #6957" (closed as completed, but not a closing-keyword link on this specific PR and not itself an authorization for this PR's own protected-path edits). Absence of a linked issue means insufficient context under this repo's protected-path policy, notwithstanding the PR's detailed description and ADR 0119. This finding has been raised and left unresolved across many prior review rounds on this PR.
    Remediation: A human reviewer must explicitly approve the changes to .github/workflows/fullsend.yaml and .github/workflows/reusable-dispatch.yml before merge, regardless of authorization status. Separately, link a tracking issue for this PR's own scope, or add explicit reasoning in the PR body for why this ADR-driven, repo-wide, opt-out run-name change is self-authorizing given the unmerged prerequisite stack (refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447 et al.).

Medium

  • [fail-open] internal/cli/steer.go:321 — FULLSEND_STEER_ACTIVE is the agent-side gate for trusting a "Runner update:" envelope: iteration.env is sourced last, and an unwatched iteration must unset it so a value planted in agent-writable .env.d cannot survive (.env sources .env.d/*.env in the middle of its script, then iteration.env last). The success path does this correctly (iterationEnvCommand writes unset FULLSEND_STEER_ACTIVE when the watcher is not live). The export-failure path in exportIterationEnv does the opposite: on writeIterationEnv failure it calls clearIterationEnv, which rm -fs iteration.env. With that file gone, iterationEnvSourceLine's if [ -f iteration.env ]; then . iteration.env; fi is a no-op, so a value an earlier iteration wrote to .env.d (export FULLSEND_STEER_ACTIVE=1) is what the next iteration inherits instead of being force-unset. The watcher is stopped on this path, so no real steer is delivered this iteration — but the flag's stated purpose is to make the agent distrust the public runner-update opening line wherever it appears, and that purpose is defeated if a stale flag survives a write failure. The function's own doc comment ("after a failure the sandbox no longer advertises it") does not match this behavior.
    Remediation: On writeIterationEnv failure, do not delete iteration.env. Write a fallback file that only unsets FULLSEND_STEER_ACTIVE (and clears the deadline if still required), then stop the watcher. Keep rm -f only if a subsequent write of the unset-only file succeeds.

Low

  • [logic-error] internal/steerwatch/delta.go:320 — bindTriggers still applies maxDispatchLag (10 minutes, run CreatedAt minus comment CreatedAt) to every pair, including pairs bound by an exact CommentID match (the comment:<id> run-name suffix this PR adds so pairing is exact rather than timestamp-based). A named pair can still be discarded as unbound if the run record's CreatedAt lags the comment by more than 10 minutes; the comment then lands as context rather than an amendment (fail-safe: the queued follow-up still does the work). pairRuns produces both exact-CommentID pairs and timestamp-window pairs into the same list, and bindTriggers applies covers() + maxDispatchLag uniformly to all of them. TestBindTriggers_DispatchLagBound covers only the unnamed-authorization path (no CommentID), so the named-id path's behavior is untested. Unchanged since prior review rounds.
    Remediation: When a pair is bound via CommentID (exact match), skip the maxDispatchLag check and keep only covers() (so an edited-after-authorization comment is still excluded), or add a named-id lag test that pins today's behavior as intentional if it is.

  • [stack-sequencing] docs/cli/steer.md:12 — ADR 0119's "Consequences" section and docs/contributing/steering.md disclose that, absent a processing receipt, an in-flight run absorbing a steer does not stop the separately-queued follow-up run from redoing the same work. docs/cli/steer.md does not carry the same caveat: it states the run in flight "absorbs it instead of being cancelled and restarted," which reads as though the follow-up run is skipped. DefaultSteerEnabled stays false, so this is an intentionally scoped, disclosed gap elsewhere; the residual concern is that the CLI's own user-facing doc is the one place this expectation isn't set. Unchanged since prior review rounds.
    Remediation: Until a receipt/suppression loop exists, keep steering off by default (already the case) and add the same caveat to docs/cli/steer.md: the queued follow-up run still executes even after a successful steer.

  • [cross-repo-env-contract] docs/guides/user/building-custom-agents.md:73 — This PR documents FULLSEND_STEER_ACTIVE in docs/cli/run.md as the variable agent definitions key off of to distinguish a legitimate runner-update message from an injection attempt. Verified directly against the PR's actual base branch (steer-provenance, not main): the runner-environment-variable lists in docs/guides/user/building-custom-agents.md ("present in every agent's shell") and the parallel section in docs/guides/user/bring-your-own-agent.md still do not mention it. Note FULLSEND_STEER_ACTIVE is set only while an iteration's watcher is actually running, not unconditionally, so any addition to these lists should carry that caveat rather than a blanket "always present" claim. Unchanged since prior review rounds.
    Remediation: Add FULLSEND_STEER_ACTIVE to the runner-environment-variable lists in docs/guides/user/building-custom-agents.md and docs/guides/user/bring-your-own-agent.md, noting it is set only while a watcher is running, with a pointer to docs/cli/run.md#run-baseline.

Resolved since the last review (methodology correction, not a code change): This round fetched internal/harness/compose.go, docs/contributing/harness-fields.md, and docs/runtimes.md directly from this PR's actual GitHub base branch (steer-provenance) rather than from target-repo/, which tracks main and does not include this feature's stack at all. Both docs/contributing/harness-fields.md (a steer row already present in the "Fields that stay at top level only" and "Merge and inheritance rules" tables) and docs/runtimes.md (a "Steer" row already present in the "Choosing a runtime" comparison table) already carry the content that many prior review rounds flagged as missing under the [config-schema-contract] / [missing-doc] findings. compose.go's mergeBaseIntoChild on steer-provenance also already merges Harness.Steer (if child.Steer == nil { child.Steer = base.Steer }). These three items are not gaps in this PR or in its unmerged prerequisite stack — they were fixed by an earlier, already-integrated commit on the real base branch, and this and prior rounds' comparisons against main produced a persistent false positive. Not re-raised this round.

Not re-raised this round:

  • The [injection-pattern] finding against internal/runtime/pi_steer.go's JSON-encoded, shell-appended mailbox write (appendLine) is not raised this round. appendLine's own implementation is still not present anywhere in this PR's diff or materialized head tree; it ships in the stacked, still-open, unmerged base PR refactor(steerwatch): verify follow-up run provenance and split the delta by authority #7447. Whether the mailbox write is shell-injectable turns entirely on a file outside this PR's own scope and cannot be confirmed or ruled out here, consistent with the most recent prior round's judgment on this point.
  • A [code-organization] finding about internal/cli/steercmd.go's file naming (command definition in steercmd.go rather than steer.go) and a [doc-style] finding about a doc comment spanning two methods in internal/cli/steer.go:552 were both raised by this round's style-conventions pass and then removed on adversarial review: internal/cli/ already has several precedents for splitting command helpers across files without a uniform *cmd.go convention, and the doc-comment issue is a cosmetic godoc adjacency between two trivial one-line wrapper methods with no contract or correctness impact.
  • Two prior-round [api-shape] and [doc-style] findings against internal/cli/steercmd.go (workItemFromSegments's unused kind return value; a doc comment referencing a nonexistent item parameter on buildSteerComment) are not observed in the current diff — workItemFromSegments now returns the narrower (owner, repo string, number int, err error) and buildSteerComment's doc comment now correctly references isPR.
  • A prior-round [code-organization] finding about ListWorkflowRunsSince's placement in internal/forge/fake.go is not observed in the current diff — it is now colocated immediately after ListWorkflowRuns, with the other workflow-run methods.

Next steps:

  • /fs-fix — agent addresses review findings automatically
  • /fs-fix <your instruction> — agent fixes with your specific guidance
  • Push commits directly — review re-runs automatically on push
  • /fs-fix-stop — disable automatic fix runs for this PR

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 6:01 PM UTC · Completed 6:31 PM UTC

Commit: 09fdce7 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $13.50

@github-actions

github-actions Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

Site preview

Preview: https://7ac0159c-site.fullsend-ai.workers.dev

Commit: 9aae704be4c8242ebcf599c4d33a86a68bfa6a69

@fullsend-ai-review fullsend-ai-review Bot added the risk/elevated PR risk: elevated label Sep 18, 2026
@fullsend-ai-review

fullsend-ai-review Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

Risk Assessment: elevated (3/5)

Details

Large diff spanning 45 files and 5,506 lines touches protected CI workflows and high-churn CLI/forge components with no linked closing issue, yielding an elevated composite risk score of 3 anchored to prior rounds; DefaultSteerEnabled remains false, bounding runtime blast radius.

Previous run

Risk Assessment: elevated (3/5)

Details

Large diff spanning 44 files and 5,402 lines touches protected CI workflows and high-churn CLI/forge components with no linked closing issue, yielding an elevated composite risk score of 3, consistent with prior rounds; DefaultSteerEnabled remains false, bounding runtime blast radius.

Previous run (2)

Risk Assessment: elevated (3/5)

Details

Large diff spanning 44 files and 5,402 lines touches protected CI workflows and high-churn CLI/forge components with no linked closing issue, yielding an elevated composite risk score of 3, consistent with prior rounds; DefaultSteerEnabled remains false, bounding runtime blast radius.

Previous run (3)

Risk Assessment: elevated (3/5)

Details

Large diff spanning 44 files and 5,362 lines touches protected CI workflows and high-churn CLI/forge components with no linked closing issue, yielding an elevated composite risk score of 3, consistent with the prior round; DefaultSteerEnabled remains false, bounding runtime blast radius.

Previous run (4)

Risk Assessment: elevated (3/5)

Details

Large diff spanning 42 files and 5,249 lines touches protected CI workflows and high-churn CLI/forge components with no linked closing issue, yielding an elevated composite risk score of 3; DefaultSteerEnabled remains false, bounding runtime blast radius.

Previous run (5)

Risk Assessment: elevated (3/5)

Details

Large diff across 42 files touching CI workflows and protected paths with high churn and author contention on core CLI files, balanced by opt-in feature gating (DefaultSteerEnabled=false) and no dependency or security-sensitive-file changes.

Previous run (6)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 signals remain elevated with large blast radius (42 files, 5,247 lines), CI workflow modifications, and 2 protected .github/ paths, compounded by high Tier 2 churn and multi-author contention on modified base files, balanced by default-disabled feature gating (DefaultSteerEnabled stays false) and scope alignment with linked issue #6957.

Previous run (7)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 metadata signals (42 files, 5,169 lines changed, large blast radius, CI workflow modifications, and 2 protected paths) remain in elevated risk buckets, combined with high Tier 2 churn, multi-author contention, and regression history across touched base files, yielding a composite score of 3 with no linked closing issue.

Previous run (8)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 metadata signals (42 files, 5,139 lines changed, large blast radius, CI workflow modifications, and 2 protected paths) remain in elevated risk buckets, combined with high Tier 2 churn, multi-author contention, and regression history across touched base files, yielding a composite score of 3 with no linked closing issue.

Previous run (9)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 metadata signals (42 files, 5,136 lines changed, large blast radius, CI workflow modifications, and 2 protected paths) remain in elevated risk buckets, combined with high Tier 2 churn, multi-author contention, and regression history across touched base files, yielding a composite score of 3 with no linked issue.

Previous run (10)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 metadata signals (42 files, 5,082 lines changed, large blast radius, CI workflow modifications, and 2 protected paths) remain in elevated risk buckets, combined with high Tier 2 churn, multi-author contention, and regression history across touched base files, yielding a composite score of 3 with no linked issue.

Previous run (11)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 metadata signals (42 files, 5,082 lines changed, large blast radius, CI workflow modifications, and 2 protected paths) remain in elevated risk buckets, combined with high Tier 2 churn, multi-author contention, and regression history across touched base files, yielding a composite score of 3 with no linked issue.

Previous run (12)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring preserves score 3 (elevated): Tier 1 metadata signals (42 files, 5,073 lines changed, large blast radius, CI workflow modifications, and 2 protected paths) remain in elevated risk buckets, combined with high Tier 2 churn, multi-author contention, and regression history across touched base files, yielding a composite score of 3 with no linked issue.

Previous run (13)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved as Tier 1 signals (42 files, 4,732 lines changed, large blast radius, and 2 protected CI workflow paths) remain in the same risk buckets alongside high Tier 2 churn, multi-author contention, and regression history on touched base files, weighted at 62% Tier 1 and 38% Tier 2 given no linked issue.

Previous run (14)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved as incremental additions (+2 files, +130 lines, totaling 42 files and 4,709 lines) remain within the same Tier 1 risk buckets alongside 2 protected CI workflow paths, with Tier 2 reflecting high churn, contention, and change coupling on touched base files, weighted at 62% Tier 1 and 38% Tier 2 given no linked issue.

Previous run (15)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved because Tier 1 metadata signals remain in the same risk tiers as the prior review (40 files, 4,579 lines, 2 protected CI workflow paths, no dependency changes, non-bot/non-first-time author) and Tier 2 git history indicates high churn, multi-author contention, and change coupling across the modified core files, with weights redistributed to Tier 1 (62%) and Tier 2 (38%) given no linked issue.

Previous run (16)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved because Tier 1 metadata signals remain in the same risk tiers as the prior review (39 files, 4,573 lines, 2 protected CI workflow paths, no dependency changes, non-bot/non-first-time author) and Tier 2 git history indicates high churn, multi-author contention, and change coupling across the modified core files, with weights redistributed to Tier 1 (62%) and Tier 2 (38%) given no linked issue.

Previous run (17)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved because Tier 1 metadata signals remain unchanged (38 files, 4,564 lines, 2 protected workflow paths, CI workflow changes, non-bot author) and Tier 2 git history indicates high churn, multi-author contention, and change coupling on pre-existing modified files, with weights redistributed to Tier 1 (62%) and Tier 2 (38%) due to no linked issue.

Previous run (18)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved because Tier 1 metadata signals remain unchanged (38 files, 4,545 lines, 2 protected workflow paths, CI workflow changes, non-bot author) and Tier 2 git history indicates high churn, author contention, and change coupling on pre-existing modified files, with weights redistributed to Tier 1 (62%) and Tier 2 (38%) due to no linked issue.

Previous run (19)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: score 3 (elevated) is preserved because Tier 1 metadata signals remain unchanged in bucket terms (38 files in 26-50 range, 4,545 lines in >=2,000 range, 2 protected workflow paths, CI workflow modified, no dependency changes, non-bot/non-first-time author) and Tier 2 git history confirms high churn and author contention on pre-existing files, with weights redistributed to Tier 1 (62%) and Tier 2 (38%) due to no linked closing issue.

Previous run (20)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: Tier 1 signals are unchanged in bucket terms from the prior assessment (38 vs 33 files both fall in the 26-50 bucket, 4425 vs 4169 lines both fall in the >=2000 bucket; still 2 protected .github/workflows/ paths, CI workflow changed, no dependency changes, non-bot/non-first-time author), and Tier 2 spot-checks on the pre-existing hot files reconfirm the same very-high churn already noted previously with no new delta. The prior composite score of 3 (elevated) is preserved with weights redistributed to Tier1=62%/Tier2=38% due to no linked issue.

Previous run (21)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring applied: Tier 1 signals are unchanged from the prior assessment (identical 33-file/4169-line diff, 2 protected .github/workflows/ paths, CI workflow change, no dependency changes, non-bot/non-first-time author), and Tier 2 spot-checks confirm the previously noted very-high churn on touched workflow/hub files with no new delta, so the prior composite score of 3 (elevated) is preserved with weights redistributed to Tier1=62%/Tier2=38% due to no linked issue.

Previous run (22)

Risk Assessment: elevated (3/5)

Details

No linked issue (weights redistributed to Tier1=62%/Tier2=38%). Tier 1 (~2.75) reflects a large 33-file/4169-line diff, two protected .github/workflows/ paths, and a CI workflow change, offset by a non-first-time/non-bot author and no dependency changes. Tier 2 (~3.6) reflects very high 30-day churn, broad multi-author contention, and heavy fix/revert history on the pre-existing forge/CLI/workflow hub files this PR touches, offset by very recent last-touch dates; new steerwatch/steer-CLI files have no independent history. Composite rounds to 3 (elevated), consistent with the prior assessment via re-review anchoring.

Previous run (23)

Risk Assessment: elevated (3/5)

Details

No linked issue (weights redistributed to Tier1=62%/Tier2=38%; PR #7447 referenced as base is a prerequisite PR, not a tracked issue, and remains open/unmerged). Tier 1 (~3.0) is driven by a large 33-file/4169-line diff, two protected .github/workflows/ paths, and a CI workflow change, offset by a non-first-time/non-bot author and no dependency changes. Tier 2 (~2.86) reproduces very high 30-day churn, broad multi-author contention, and heavy fix/revert history on the pre-existing forge/CLI hub files this PR touches, offset by very recent last-touch dates. Composite rounds to 3 (elevated), consistent with the prior assessment via re-review anchoring; file count and line count match the prior round exactly.

Previous run (24)

Risk Assessment: elevated (3/5)

Details

No linked issue (weights redistributed to Tier1=62%/Tier2=38%); Tier 1 (~2.75) is driven by a large 33-file/4169-line diff, two protected .github/workflows/ paths, and a CI workflow change, offset by a non-first-time/non-bot author and no dependency changes; Tier 2 (~2.86) reflects very high historical churn, multi-author contention, and fix/revert frequency on pre-existing forge/CLI hub files this PR touches, while new steerwatch/steer-CLI code is excluded under the new-file rule. Composite rounds to 3 (elevated), consistent with the prior assessment via re-review anchoring; the prior rationale's claim that sibling PR #7447 was merged is corrected as inaccurate this round (it remains open) but is not load-bearing for the score.

Previous run (25)

Risk Assessment: elevated (3/5)

Details

Tier 1 signals (33 files, 4162 lines, large blast radius, 2 protected workflow paths, CI change, no dependency changes, non-bot/non-first-time author) fall in the same buckets as the prior round, and Tier 2 confirms the same high-churn/regression-fix pattern on pre-existing hub files while new steering logic remains excluded under the new-file rule; the merge of sibling PR #7447 into the base branch does not change Tier 3 availability (still no linked issue), so the prior composite of 3 (elevated) is preserved via re-review anchoring.

Previous run (26)

Risk Assessment: elevated (3/5)

Details

No linked tracked issue, so weights redistribute to Tier1=62%/Tier2=38%. Tier1 (~2.75) driven by large diff (33 files/4022 lines, blast=large), two protected .github/workflows/ paths, and a CI workflow change, offset by a non-first-time/non-bot author and no dependency changes. Tier2 reflects high historical churn/regression-fix rates on pre-existing hub files this PR touches (e.g. internal/cli/run.go), while the bulk of new steering logic is excluded from Tier2 under the new-file rule. Composite rounds to 3 (elevated), consistent with the prior assessment across multiple rounds on this PR.

Previous run (27)

Risk Assessment: elevated (3/5)

Details

No linked issue, so weights redistribute to Tier1=62%/Tier2=38%; Tier1 (~2.75) is driven by large diff size (33 files/3892 lines, blast=large), two protected .github/workflows/ paths, and a CI workflow change, offset by a non-first-time/non-bot author and no dependency changes; Tier2 (~2.71) reflects high historical churn/regression-fix rates and broad change-coupling on pre-existing hub files this PR touches, even though most new code is brand-new and excluded from Tier2 by the new-file rule; composite rounds to 3 (elevated), consistent with the prior assessment on this PR.

Previous run (28)

Risk Assessment: elevated (3/5)

Details

Re-review anchoring: Tier 1 signals unchanged bucket-for-bucket from the prior assessment (31 files, 2 protected .github/ paths, CI workflow changed, no dependency changes, non-bot/non-first-time author); Tier 2 confirms brand-new steer/steerwatch code plus continued high-churn edits to a small set of shared infrastructure files. No new risk factor since the prior round; composite preserved at 3 (elevated).

Previous run (29)

Risk Assessment: elevated (3/5)

Details

Large multi-part feature PR (31 files, 3635 lines, 2 protected .github/ paths, CI workflow changed) drives an elevated Tier 1 composite (~2.75), moderated by a non-bot, non-first-time author and no dependency changes; Tier 2 is dominated by mostly brand-new files but includes real coupling risk from editing internal/cli/run.go, a very high-churn, multi-author, fix-prone core file. Weighted composite (62% Tier 1 / 38% Tier 2, no linked issue) rounds to 3 (elevated), consistent with a refactor-of-widely-touched-module-with-high-churn-and-no-linked-issue risk band.

fullsend-ai-review[bot]

This comment was marked as outdated.

fullsend-ai-review[bot]

This comment was marked as outdated.

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 9:36 PM UTC · Completed 9:59 PM UTC

Commit: a3e6ae9 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $12.05

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 10:07 PM UTC · Completed 10:30 PM UTC

Commit: dfb62b8 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $11.45

fullsend-ai-review[bot]

This comment was marked as outdated.

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 11:18 PM UTC · Completed 11:47 PM UTC

Commit: f326c4b · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $17.09

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 12:23 AM UTC · Completed 12:52 AM UTC

Commit: bb2f424 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $15.09

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ❌ Failure (agent timed out after 45m0s without completing (timeout: 45m0s)) · Started 4:06 PM UTC · Completed 4:52 PM UTC

Commit: 3718ca9 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $14.10

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 5:09 PM UTC · Completed 5:41 PM UTC

Commit: 3718ca9 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $19.08

fullsend-ai-review[bot]

This comment was marked as outdated.

@waynesun09

Copy link
Copy Markdown
Member Author

On the two [stale-doc] Mediums in the latest review (3718ca9): both files already carry the steer content at this head. It arrives through the lower PRs in this stack, so neither file is in this PR's own diff.

  • docs/contributing/harness-fields.md has steer in the top-level-only table (L65) and in the merge-rules table (L98).
  • docs/runtimes.md has the Steer row (L67) inside "Choosing a runtime", which is the section harness-reference.md links to.

No change needed in this PR.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 7:16 PM UTC · Completed 7:51 PM UTC

Commit: 90251ba · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $16.92

fullsend-ai-review[bot]

This comment was marked as outdated.

@fullsend-ai-review

fullsend-ai-review Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ❌ Failure (validation failed after 1 iteration(s)) · Started 8:28 PM UTC · Completed 8:41 PM UTC

Commit: e161ab4 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $6.82

@fullsend-ai-review

Copy link
Copy Markdown

🤖 Review · Started 8:56 PM UTC

Commit: e161ab4 · View workflow run →

…sses

The watcher had the provenance checks but nothing driving them. Add the
loop: poll on a ticker and immediately at every turn end, fold every
candidate a poll accepts into one steer, and settle when nothing new
arrived, the cap is spent, the deadline passes or the context is done.

The listing re-derives from the run's own start on every poll rather than
advancing a cursor, so a candidate is bounded by freshness and deduplicated
by run id. ListWorkflowRunsSince filters server-side to one workflow file
and paginates to a fixed cap; GitLab returns ErrNotSupported, which is what
the runner's GitHub-only gate already expects.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
…dentity

Wire the watcher into the run: one per validation-loop iteration, carrying
the judged run ids, delta window and steer count across iterations so
max_steers stays a cap on the run. A steered run is bounded by a whole-run
deadline anchored at the run's start rather than a per-iteration timeout,
and the sandbox, the heartbeat and the timeout detection all read that same
pair so none of them can promise time the run will not get.

Resolve the logins this run's own output is posted under before bootstrap,
and decline steering as an announced environment defect when they cannot be
resolved. Steering without them would let a run absorb its own post-fix and
post-code comments, which carry no marker to recognise them by, so no
steering is the correct answer rather than a shorter exclusion list.

Export FULLSEND_STEER_ACTIVE only for an iteration whose watcher is already
running. The agent definitions treat the runner-update opening line as an
injection attempt unless the variable is set, and that line is public, so
exporting it before the watcher's live calls have succeeded would tell the
agent to trust text any author can write.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
…workstation

issue_comment and issues runs carry no pull_requests[], so an in-flight run
has no server-side field to tell a follow-up run on its own work item from
one on another. Give the per-repo shim a run-name naming the item, which the
Actions API returns as display_title, and pin it against drift. A
comment-triggered run appends its comment id, so a command pairs to its run
exactly rather than by a timestamp that can cross when runs queue out of
order.

`fullsend steer <url> "<text>"` posts the stage's own slash command rather
than a command of its own: the existing route arms already carry the right
authorization floor for each stage, and the watcher accepts a follow-up run
on provenance rather than on which words the comment opened with, so an
ordinary /fs-review is absorbed exactly as anything else would be.

BREAKING CHANGE: the per-repo shim's run-name retitles every Actions run in
a repository that syncs it. Run titles now read `<owner>/<repo>#<number>`
instead of GitHub's per-event default, and a comment-triggered run adds a
` comment:<id>` suffix, as in `acme/widgets#7 comment:2551234567`. The
change arrives through scaffold sync rather than on opt-in — it is not gated
behind steer.enabled. Update any dashboards, saved filters or scripts that
match runs by title.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
ADR 0106 rejected polling forge events inside `fullsend run` because that
needs input drivers, race-safe cursors and scheduling in the execution
command. Record the mechanism that is left — the runner lists the execution
platform's own run records for its own shim, with the job token — and the
invariants that keep ADR 0106's limits intact.

Document the mechanics beside the interface they extend: how a candidate is
discovered and bound to this run, how a steered run settles, and the token,
cost and remaining-time ceilings it runs inside. `max_steers` and
`poll_interval_seconds` become live here; until now nothing read them.

FULLSEND_STEER_ACTIVE is documented with the other run facts: the agents key
the runner-update opening line on it, so absence is the default and presence
means a watcher is running right now.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
…tarts

Pi steers only a run that cannot fall back across models, and until now
learned that inside Run: the watcher had already started, polled, and
spent a steer slot before Steer reported the run unsupported.

Add runtime.SteerDecliner. The runner asks SteerDeclineReason in
steerEligible, before the watcher starts, and announces the decline
whatever the harness says: it follows from the run's configuration, and
with steering on by default no harness names it explicitly.

pi_steer.go: gate input shared with eligibility, behaviour unchanged.
piSteerGate and SteerDeclineReason both ask piSteerDecline over the
chain piRunChain derives, which Run now uses too, so the two cannot
disagree.

run.go builds RunParams once, before the watcher, and passes the same
value to the decline check and to rt.Run.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
The pull request branch of buildDelta returned after the head and the
reviews, and the label diff ran only for issues. An accepted `labeled`
follow-up on a pull request therefore built an empty delta: the run was
judged and marked seen but never delivered, so the update reached the
agent only if something else changed. That is the shape of a pull
request the coder App opens and labels ready-for-review a second later.

Start now snapshots a pull request's labels from the issue record GitHub
keeps for it, and every poll diffs them the way the issue path does,
through a shared labelChange. The label baseline advances with a
delivered steer, as an issue's does, so the same label is not sent twice.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
@fullsend-ai-review

fullsend-ai-review Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 9:10 PM UTC · Completed 9:41 PM UTC

Commit: 9aae704 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5 · Effort: high · Cost: $15.01

@fullsend-ai-review fullsend-ai-review Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

See the review comment for full details.

return fmt.Errorf("reading labels of %s#%d: %w", w.cfg.Repo, w.cfg.Item.Number, err)
}
if issue != nil {
w.cfg.Item.Labels = issue.Labels

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[low] logic-error

resolveItem() snapshots live labels/title/body at Start() (after checkout/mint/sandbox bootstrap, immediately before the agent starts). This is the intended baseline: resolveItem's own comment explains that an empty snapshot would report every already-present label as newly added on every poll. Comment-based deltas are windowed from freshAfter; labels are a set diff against the Start()-time snapshot instead — a deliberate difference, not an oversight, since the agent's own first turn already sees the current label set as initial context. The residual gap is narrow: a label that lands in the run's own bootstrap gap (before Start()) is already in the snapshot, so an accepted, later labeled follow-up run builds an empty delta and pollAndSteer's markSeen-without-receipt path means the queued run behind it still redoes the work per this PR's documented "no receipt yet" model. The new tests (TestBuildDelta_PullRequestLabels, TestPollAndSteer_PullRequestLabelIsDelivered) cover only a label change that happens after Start() (the actual steer path this PR targets); a label present before Start() is untested, and the PR body's own motivating example (a label added "a second after opening") sits right at this boundary.

Suggested fix: If a label landing in the run's own bootstrap gap should be treated as absorbed, add a test that sets the label on the stub before constructing/Start()-ing the watcher and assert a non-empty delta (or document explicitly that it is expected to fall to the queued run, matching the current behavior, and assert that contract instead).

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed, the pre-Start label case should be pinned, not only explained. A test is prepared: TestPollAndSteer_LabelBeforeStartIsBaseline sets the label on the item before the watcher starts and asserts the expected, documented behaviour:

  • the labeled follow-up is judged and no steer is delivered, because the label is already in the baseline the agent reads;
  • nothing is receipted, so its own queued run still does the work.

It lands with the next push of this PR.

// steer this one, so title binding is turned off: pull-request events
// still bind on pull_requests[], and an issue_comment follow-up falls
// to its own queued run.
if w.cfg.RunName != "" && runNameItem(run.DisplayTitle) != w.cfg.RunName {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[low] fail-open

The prior medium-severity finding here (steerWatcherConfig in internal/cli/steer.go always setting Config.RunName with no proof the repository's shim actually declares run-name) is fixed at this head: Watcher.Start() now clears cfg.RunName when runNameItem(run.DisplayTitle) != w.cfg.RunName, covered by TestStart_TitleBindingNeedsARunNameShim. A narrower residual remains: the check infers shim presence solely from the work-item half of DisplayTitle matching the configured RunName. Without a declared run-name, GitHub's display_title defaults to the issue/PR's own title — so a collaborator who can edit that title to exactly <owner>/<repo>#<N> (or with a comment:<id> suffix) before the victim run is created would still leave title binding enabled, letting a Route-authorized same-stage follow-up on a different item with the same crafted title false-bind. This PR's own shim template (shim-per-repo.yaml) and this repository's own shim now both declare run-name, so the residual hole applies only to repositories/shims that have not yet adopted it. Exploitation still requires an actor who can obtain an accepted, Route-authorized follow-up run — not anonymous external injection — and DefaultSteerEnabled remains false in this PR.

Suggested fix: After loading myRun in Start(), additionally treat myRun.DisplayTitle equaling the live issue/PR title (fetched independently, not just compared to the configured RunName string) as proof that no run-name shim is in effect, and clear cfg.RunName in that case too. Add a test where DisplayTitle is owner/repo#N specifically because that is the item's actual title (not because a shim set it), and assert a same-titled follow-up is rejected.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed; the residual is real. Without a run-name, GitHub titles my run with the item's own title, so an item retitled to <owner>/<repo>#<N> passed the first check.

The fix is prepared. After it reads the item, Start also turns title binding off when my run's title is exactly the item's live title. That is the forge's default title, whatever it spells, so it says nothing about the shim. A run-name shim titles the run RunName, which an item title matches only if someone set it that way, and failing closed is right then too.

TestStart_TitleBindingNeedsARunNameShim gains two cases:

  • an item titled org/repo#7: title binding is off, and only the pull-request event binds;
  • a run-name shim with an ordinary item title: title binding stays on.

Removing the new check fails the first case. One narrow gap remains: a title set before my run is created and changed back before Start reads it. The fix lands with the next push of this PR.

@waynesun09

Copy link
Copy Markdown
Member Author

Superseded by #9029; the steer design was reworked on top of ADRs 0133/0141/0144/0151.

This branch was successfully deployed

1 active and 1 inactive deployments
site-preview — 9aae704b Deployed Sep 30, 2026 by github-actions[bot]
dev — 9aae704b Deployed Sep 30, 2026 by waynesun09 via behaviour #14716
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

component/ci CI pipelines and checks component/harness Agent harness, config, and skills loading fullsend-no-fix Skip bot-triggered fix agent runs risk/elevated PR risk: elevated type/adr type/feature New capability request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants