Skip to content

docs(adr): let agent-spec sub-agent entries source or disable personas - #8969

Open
fullsend-ai-coder[bot] wants to merge 13 commits into
mainfrom
agent/8968-adr-subagent-source-disable
Open

fullsend-ai-coder[bot] wants to merge 13 commits into
mainfrom
agent/8968-adr-subagent-source-disable

Conversation

@fullsend-ai-coder

@fullsend-ai-coder fullsend-ai-coder Bot commented Oct 31, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Adds ADR 0155. It lets an agents[].subagents entry in .fullsend/config.yaml be an object {model, source, enabled} as well as today's model string. With it, a repository can replace, add or disable a review persona without forking the orchestrator skill. This PR is the decision record only; implementation (config parsing, staging, CLI flags) follows in separate work.

Related Issue

Issue #8968 (closing reference appended automatically).

Changes

  • New docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md (Accepted):
    • An entry is a string (shorthand for {model}) or {model, source, enabled}; ADR 0104's model resolution is unchanged.
    • source and enabled: false compile into harness skill file overrides at harness composition (shared with fullsend lock) and are applied at staging before persona discovery, on every runtime.
    • source and enabled: false are read only from config at the trusted ref, never from overlays, flags, env or the change under review; source is a path inside .fullsend/ or a #sha256=-pinned URL on allowed_remote_resources with a lock entry.
    • enabled: false guarantees the persona is not staged; the post-script finding drop on the agent-reported dimension is a backstop.
    • Unknown keys: still an error unless enabled: false (ignored with a warning) or source defines the persona.
    • Four options (registry file, per-dimension env flags, harness overrides only, object form), a YAML example, forge scope.
  • docs/architecture.md: adds a "Decided, not yet implemented" bullet under Agent Runtime › Decided (implementation).
  • docs/problems/code-review.md: adds a link to the ADR under "Review sub-agent decomposition".
  • docs/reference/config-reference.md: adds one line to the subagents field saying the object form is decided in ADR 0155 but not yet implemented.

ADR 0104 and every other accepted ADR are untouched. There are no Go, script, skill or workflow changes.

Testing

  • make lint equivalent: pre-commit could not fetch its remote hook repos in the sandbox (TLS certificate error), so the applicable hooks were run directly on the staged files. lint-adr-status, lint-adr-numbers, lint-adr-frontmatter, lint-docs-links, lint-broken-symlinks, lychee (offline, with fragments), gitleaks, and the trailing-whitespace / EOF / line-ending / merge-conflict checks all pass. The post-script runs the authoritative pre-commit.
  • Tests added/updated for new or modified logic: not applicable (docs only).

Notes for reviewers

  • Current-behaviour evidence (kept out of the ADR text on purpose, so the record states the problem in plain words) at 12a09ace3e26b1e52442ee628ac570f794f856f6: AgentEntry.Subagents (map[string]*string) in internal/config/config.go; SkillEntry.Overrides in internal/harness/skill_entry.go, which Bootstrap does not apply (it passes only SkillSources(h.Skills)); the unknown-persona error in resolvePersonaModels (internal/runtime/pi_personas.go) and resolveCodexChildModels (internal/runtime/codex_personas.go).

  • ADR number: open PR docs(adr): let agent entries carry harness overrides #8965 claims 0154. No open PR claims 0155. If 0155 is taken before this merges, use the renumber-adr skill.

Checklist

  • PR title follows Conventional Commits
  • Commits are not signed off (autonomous agent commit, DCO-exempt)

🤖 Generated with Claude Code


Closes #8968

Post-script verification

  • Branch is not main/master (agent/8968-adr-subagent-source-disable)
  • Secret scan passed (gitleaks — 12a09ace3e26b1e52442ee628ac570f794f856f6..HEAD)
  • PR body secret scan passed (gitleaks — no-git)
  • Final PR description secret scan passed (gitleaks — no-git)

@waynesun09

Copy link
Copy Markdown
Member

/fs-fix-stop

@github-actions github-actions Bot added the fullsend-no-fix Skip bot-triggered fix agent runs label Oct 31, 2026
@github-actions

Copy link
Copy Markdown

Fix agent disabled for this PR. Remove the fullsend-no-fix label or use /fs-fix to re-engage.

@fullsend-ai-review

fullsend-ai-review Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 4:01 PM UTC · Completed 4:10 PM UTC

Commit: 688fbe1 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $1.64

@github-actions

github-actions Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

Site preview

Preview: https://c25ce5db-site.fullsend-ai.workers.dev

Commit: 4c99d51b14805e9a8c8dde5b020e8e8910e8ad9e

@waynesun09

Copy link
Copy Markdown
Member

ADR 0155 fix round 1

Apply all items below to docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md (and the living docs where noted) in one commit. Do not sign off: no git commit -s, no Signed-off-by trailer.

Style rule for the whole ADR: state the problem and the decision in plain words. Context and Decision must contain no commit hash, no file path and no Go identifier (no AgentEntry, SkillEntry, map[string]*string, resolvePersonaModels, .go files). Keep the existing Status, the rejected options, the cross-links to ADRs 0104, 0112 and 0018, and the Options section.

Context

  1. Replace paragraph 1 with:

    Each agent entry's subagents map sets a model per persona (ADR 0104). It cannot point a persona at a different file or turn one off. Harness skills: entries can replace or remove files inside a pinned skill, but staging does not apply those overrides yet (#6158). So today the only way to add, replace or remove a review persona is to copy the whole skill, and every upstream change to the skill must then be ported by hand. Copying does not make selection deterministic either: the orchestrator still chooses among the copied personas.

    Keep the code-review.md link from the old paragraph if it fits. Mention feat(harness): file-level overrides within skill directories #6158 only here.

Decision

  1. Item 1: add "A string value is shorthand for {model: <value>}." The reserved default key accepts only the string form.
  2. Item 2: say "the harness skill file overrides" (no issue number). Say the overrides are applied before persona discovery, so discovery and model resolution see the staged tree. Say where a new persona goes: a source for a persona that no staged skill ships adds sub-agents/<persona>.md to the agent's first skill that already ships personas. The sourced file's frontmatter name must equal the key. Remove the repeated "add".
  3. Item 3: do not call source "a guarded field in the sense of ADR 0112". Say instead that source crosses a trust boundary, so, like ADR 0112's guarded fields, it is accepted only from config at the trusted ref (the same ref config.yaml and the harness are read from) and never from an overlay, a CLI flag, an environment variable or the change under review. A repository path resolves against the same .fullsend/ directory as harness skill overrides and must not leave it (no .., no absolute path, no symlink escape). A URL needs #sha256=, a prefix on allowed_remote_resources, and a lock entry: fullsend lock runs the same compile step as the run, and a run fails on an unlocked or mismatched source. Replace "read from the base branch" with "read from the trusted ref".
  4. Item 4: keep both effects, but word the second honestly. The guarantee is that the persona file is not staged, so no child can be dispatched as that persona. Then: "As a backstop, findings record the persona that produced them in a dimension field that the review skill fills in, and the review post-script drops findings whose dimension names a persona that was not staged. The field is reported by the agent, so this filters mislabelled output and is not proof of what ran." Keep the sentence about the parent doing the work itself.
  5. Item 5 (layering): after "merge per field", add: "ADR 0104's rule that a shared preset must not carry subagents entries still holds for model and source entries until the merge records which layer a key came from."
  6. Item 8: remove the function and file names. Plain form: "Today a subagents key that names a persona the run did not discover is an error." Then state three cases:
    • a model-only entry that names no persona in the staged tree stays an error;
    • an entry with source defines the persona, so it is never undiscovered;
    • an enabled: false entry for a persona that no staged skill ships is ignored with a warning in the run log naming the discovered personas. An org-wide disable then still works in repositories without the persona, and a typo shows up.
  7. New short item (runtimes): "source and enabled act at staging, so they apply on every runtime. model applies where ADR 0104's resolution is implemented (pi, and codex when the agent delegates); elsewhere it is ignored as today."
  8. YAML example:
    • use style-conventions: (a real persona name), not style:;
    • use the raw-content URL form https://raw.githubusercontent.com/example-org/personas/<sha>/licensing.md#sha256=<sha256>;
    • keep the comments consistent with item 8.

Forge scope

  1. Forgejo has no forge implementation yet, so do not say it works the same way. Write: "Forge-neutral: config, staging and the post-script do not depend on the forge. It applies on GitHub and GitLab, each with its own trusted ref, and on Forgejo once that forge lands."

Consequences (keep 3–5 one-sentence bullets)

  1. Replace the CLI bullet with "The CLI gains a way to set source and to disable a persona, implemented after this ADR is approved." Use no flag names.
  2. Replace the last bullet with "Disabling or adding a persona takes effect only once the review skill builds its roster from the staged persona files and fills in dimension, which is agents-repository work that must ship first."
  3. Bullet 2: align with item 5, so that "not staged" is the host guarantee and the finding drop is a backstop.

Living docs

  1. In the docs/architecture.md bullet, align the finding-drop wording with item 5, and change "read from the base branch" to "read from the trusted ref".

Final check

  1. Grep the ADR's Context and Decision for any remaining hash, file path or Go identifier, and remove what you find.

@waynesun09

Copy link
Copy Markdown
Member

/fs-fix Apply every item in the "ADR 0155 fix round 1" comment: #8969 (comment)
Do not sign off: no git commit -s, no Signed-off-by trailer.

@fullsend-ai-coder

fullsend-ai-coder Bot commented Oct 31, 2026 •

Copy link
Copy Markdown
Contributor Author

🤖 Finished Fix · ✅ Success · Started 4:10 PM UTC · Completed 4:14 PM UTC

Commit: 688fbe1 · View workflow run →

Runtime: claude · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $0.36

@fullsend-ai-review fullsend-ai-review Bot added the risk/moderate PR risk: moderate label Oct 31, 2026
@fullsend-ai-review

fullsend-ai-review Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

Risk Assessment: moderate (2/5)

Details

Documentation-only PR by a bot with complete issue scope alignment, elevated to moderate risk due to high historical churn, multi-author contention, and co-change coupling across touched architecture and reference documents.

Previous run

Risk Assessment: moderate (2/5)

Details

Documentation-only PR by a bot with no protected paths or code changes, elevated to moderate risk by high churn and multi-author contention on central living documentation files.

Previous run (2)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only ADR by a bot account with no protected paths or code changes, elevated to moderate risk due to high recent commit churn and author contention across touched living architecture and reference docs.

Previous run (3)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only ADR by a bot account with no protected paths or code changes, elevated to moderate risk due to high recent commit churn and author contention across touched living architecture and reference docs.

Previous run (4)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only ADR by a bot account with no protected paths or code changes, elevated to moderate risk due to high recent commit churn and author contention across touched living architecture and reference docs.

Previous run (5)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only ADR by a bot account with no protected paths or code changes, elevated to moderate risk due to high recent commit churn and author contention across touched living architecture and reference docs.

Previous run (6)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only ADR by a bot account with no protected paths or code changes, elevated to moderate risk due to high recent commit churn and author contention across touched living architecture and reference docs.

Previous run (7)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only PR by a bot with complete issue scope alignment, elevated to moderate risk due to high historical churn and co-change coupling on touched architecture and reference documents.

Previous run (8)

<!-- fullsend:risk-assessment -->
Risk Assessment: moderate (2/5)

Details

Documentation-only PR by a bot with complete issue scope alignment, elevated to moderate risk due to high historical churn and co-change coupling on touched architecture and reference documents.

@fullsend-ai-review

fullsend-ai-review Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

Review

Findings

Medium

  • [pattern-violation] docs/ADRs/0155-subagent-entries-may-source-or-disable-personas.md:2 — ADR number 0155 is already taken on current main by docs/ADRs/0155-configure-vault-backed-runner-secrets.md ("155. Select native or Vault sources for runner secrets"). Adding a second 0155 record breaks the unique-number check in hack/lint-adr-numbers, so lint will fail once this is merged with main. Any reference to "ADR 0155" would also point to two different decisions.
    Remediation: Pick an unused ADR number (0154 and 0156 are also taken on main), rename the new file, and update its frontmatter title and H1. Then update every link and ADR-number reference this PR adds in architecture.md, harness-fields.md, problems/code-review.md and config-reference.md. Leave the existing accepted Vault ADR unchanged.

Earlier findings

  • docs/contributing/harness-fields.md:14, incomplete-doc, resolved — the latest push adds an ADR 0155 paragraph to the "Decided, not yet implemented" callout in that file.

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

<!-- fullsend:review-findings-v2:eyJ2ZXJzaW9uIjoyLCJhY3Rpb24iOiJyZXF1ZXN0LWNoYW5nZXMiLCJmaW5kaW5ncyI6W3sic2V2ZXJpdHkiOiJtZWRpdW0iLCJjYXRlZ29yeSI6ImluY29tcGxldGUtZG9jIiwiZmlsZSI6ImRvY3MvY29udHJpYnV0aW5nL2hhcm5lc3MtZmllbGRzLm1kIiwiaWQiOiJmXzkyYTA3ZTVhYmY4MzZkMTAiLCJsaW5lIjoxNCwiYWN0aW9uYWJsZSI6dHJ1ZSwiZGltZW5zaW9uIjoiZG9jcy1jdXJyZW5jeSJ9XX0= -->

Review

Findings

Medium

  • [incomplete-doc] docs/contributing/harness-fields.md:14 — ADR 0155 decides that subagents becomes a harness field (set as agents[].harness.subagents or in a custom harness file) that compiles into skills: file overrides, with overlay rules (subagents.*.source and subagents.*.enabled cannot be set by any overlay; subagents.*.model and subagents.default are unguarded model settings). The ADR links to this living reference, but the "Decided, not yet implemented" callout in docs/contributing/harness-fields.md lists ADRs 0112, 0158, 0151 and 0162 and does not mention ADR 0155 or the planned subagents harness field. This file is not changed by the PR.
    Remediation: Add an entry to the "Decided, not yet implemented" callout in docs/contributing/harness-fields.md summarizing ADR 0155: subagents becomes a harness field whose source and enabled entries cannot be set in overlays, while model and default remain model settings.

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

Reason: tool-failure

This PR was NOT reviewed. Do not count this as an approval.

Previous run (3)

<!-- fullsend:review-findings-v2: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 -->

Looks good to me

Previous run (4)

<!-- fullsend:review-findings-v2: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 -->

Looks good to me

Previous run (5)

<!-- fullsend:review-findings-v2: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 -->

Review

Findings

Medium

  • [incorrect-doc] docs/architecture.md:315 — The new "Decided (implementation)" bullet still says "A disabled persona is not staged, so no child can be dispatched as it." The latest revision of ADR 0155 decision 7 removed that claim. It now says the persona prompt never reaches the sandbox and a child started under the persona name runs only as an anonymous helper. The architecture summary therefore states a stronger disable guarantee than the ADR it links to.
    Remediation: Reword lines 315–316 of docs/architecture.md to match ADR 0155 decision 7, for example: "A disabled persona is not staged, so its prompt never reaches the sandbox; a child started under its name runs only as an anonymous helper; as a backstop, ..."

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)

<!-- fullsend:review-findings-v2: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 -->

Review

Findings

Medium

  • [edge-case] docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md:118 — Decision item 7 says a disabled persona "is not staged, so no child can be dispatched as it". Removing the persona file does not guarantee that. In internal/runtime/pi_extension/fullsend-agent.js (about lines 1105–1117), an unknown non-empty subagent_type is rejected only when at least one persona is registered. With none registered, it is accepted as an anonymous child, for forward compatibility (ADR 0104). A file removed before staging also never enters skippedPersonas, so that rejection path does not apply. If configuration disables every persona the skill ships, a dispatch under a disabled name still runs. The post-script finding drop only filters output after the child has run.
    Remediation: Say that disabled persona names are kept in a host-owned dispatch deny-list and rejected even when the registered set is empty, or require strict dispatch whenever any persona is disabled. Add the all-personas-disabled case to the implementation acceptance criteria. Otherwise, soften "host guarantee" to match the mechanism described.

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)

<!-- fullsend:review-findings-v2:eyJ2ZXJzaW9uIjoyLCJhY3Rpb24iOiJyZXF1ZXN0LWNoYW5nZXMiLCJmaW5kaW5ncyI6W3sic2V2ZXJpdHkiOiJtZWRpdW0iLCJjYXRlZ29yeSI6ImRvYy1zdHlsZSIsImZpbGUiOiJkb2NzL0FEUnMvMDE1NS1hZ2VudC1zcGVjLXN1YmFnZW50cy1zb3VyY2Utb3ItZGlzYWJsZS1wZXJzb25hcy5tZCIsImlkIjoiZl9mNmU4ODA2NmU5YzQ4NzkyIiwibGluZSI6MzAsImFjdGlvbmFibGUiOnRydWV9LHsic2V2ZXJpdHkiOiJsb3ciLCJjYXRlZ29yeSI6ImRlc2lnbi1kaXJlY3Rpb24iLCJmaWxlIjoiZG9jcy9BRFJzLzAxNTUtYWdlbnQtc3BlYy1zdWJhZ2VudHMtc291cmNlLW9yLWRpc2FibGUtcGVyc29uYXMubWQiLCJpZCI6ImZfZDUwNWU1YTFlZjQxNWJmYSIsImxpbmUiOjE0NiwiYWN0aW9uYWJsZSI6ZmFsc2V9XX0= -->

Review

Findings

Medium

  • [doc-style] docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md:30 — The ADR's Context describes current behavior (the subagents map sets only a model per persona, harness skills: overrides are not applied at staging, and an unknown persona key is an error per Decision item 8) but cites no implementing source file and no commit SHA; it cites only ADR 0104 and feat(harness): file-level overrides within skill directories #6158. docs/contributing/adrs.md (lines 27-28) and skills/writing-adrs/SKILL.md (item 6) require new ADRs to cite the implementing file or function at a commit SHA, and issue docs(adr): let agent-spec sub-agent entries source or disable personas #8968 asked for citations at commit 12a09ace3e26b1e52442ee628ac570f794f856f6 by file path in plain text. The PR body says the ADR cites these, but the final text does not.
    Remediation: In Context and/or Decision item 8, cite the implementing sources in plain text at commit 12a09ace3e26b1e52442ee628ac570f794f856f6: internal/config/config.go (AgentEntry.Subagents), internal/harness/skill_entry.go (SkillEntry.Overrides), internal/runtime/pi_personas.go (resolvePersonaModels) and internal/runtime/codex_personas.go (resolveCodexChildModels), after confirming each matches the code at that SHA.

Low

  • [design-direction] docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md:146 — The Consequences bullet says only that "The CLI gains a way to set source and to disable a persona". Issue docs(adr): let agent-spec sub-agent entries source or disable personas #8968 names the surface (fullsend agent set --subagent-source and --subagent-disable), so the ADR is less specific than the issue asked for. Keeping the CLI generic is defensible for a decision record, so this is advisory.
    Remediation: Name the command and flags from the issue in the Consequences bullet, or state that the exact CLI surface is left to the implementation work.

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 (8)

Review

Reason: stale-head

The review agent reviewed commit 1a4993aa6175367306380144164b1d7065cea94f but the PR HEAD is now 20a645f97fc068bd2513b052043be2c6adeaaf67. This review was discarded to avoid approving unreviewed code.

Previous run (9)

<!-- fullsend:review-findings-v2:eyJ2ZXJzaW9uIjoyLCJhY3Rpb24iOiJyZXF1ZXN0LWNoYW5nZXMiLCJmaW5kaW5ncyI6W3sic2V2ZXJpdHkiOiJtZWRpdW0iLCJjYXRlZ29yeSI6ImxvZ2ljLWVycm9yIiwiZmlsZSI6ImRvY3MvQURScy8wMTU1LWFnZW50LXNwZWMtc3ViYWdlbnRzLXNvdXJjZS1vci1kaXNhYmxlLXBlcnNvbmFzLm1kIiwiaWQiOiJmXzc3Nzk5OTRhYzAyZTQyMWQiLCJsaW5lIjoxMTcsImFjdGlvbmFibGUiOnRydWV9LHsic2V2ZXJpdHkiOiJsb3ciLCJjYXRlZ29yeSI6ImFwaS1jb250cmFjdCIsImZpbGUiOiJkb2NzL0FEUnMvMDE1NS1hZ2VudC1zcGVjLXN1YmFnZW50cy1zb3VyY2Utb3ItZGlzYWJsZS1wZXJzb25hcy5tZCIsImlkIjoiZl9kYzM3MWQ1MWJjYzNmZjJmIiwibGluZSI6MTAwLCJhY3Rpb25hYmxlIjpmYWxzZX1dfQ== -->

Review

Findings

Medium

  • [logic-error] docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md:117 — Decision 8 keeps the undiscovered-persona error fatal for any entry that sets model or source, but makes enabled: false on an absent persona a no-op. The rules overlap for an entry with enabled: false plus model/source. Decision 5's per-field layering produces such entries automatically (org sets model, repo sets enabled: false). Decision 2 removes the disabled persona's file before discovery, so the persona is undiscovered while its entry still sets model. Today's resolvers treat that as fatal (internal/runtime/pi_personas.go and internal/runtime/codex_personas.go). The ADR's main use case, disabling a persona that has an inherited model, can therefore resolve to an error.
    Remediation: State the precedence explicitly. For example, enabled: false suppresses staging and model resolution, and any retained model/source is inert. Make Decisions 2, 5 and 8 agree with that rule.

Low

  • [api-contract] docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md:100 — Decision 4 has the review post-script drop findings "whose dimension names a persona that was not staged", but the ADR never defines which finding field carries the persona or dimension, or what happens when it is missing or unknown. Review findings today carry category and no persona field, and a category does not map to a persona name (especially for added personas such as licensing). The "findings are dropped" guarantee has no defined producer-to-consumer contract yet.
    Remediation: Say that findings must carry a persona attribution field, how the orchestrator fills it in, and how the post-script treats missing, unknown and orchestrator-owned attribution. Alternatively, list the result-schema and skill-contract change under Consequences as an implementation prerequisite.

Labels: Docs-only ADR PR


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 added type/adr component/docs User-facing documentation labels Oct 31, 2026
fullsend-ai-coder Bot added a commit that referenced this pull request Oct 31, 2026
Rewrite ADR 0155 Context and Decision in plain words (no hashes, file
paths or Go identifiers), tighten the source trust rules, reword the
enabled:false guarantee with the dimension-drop backstop, relax the
undiscovered-persona error per case, add runtimes and forge-scope
clarifications, and align docs/architecture.md.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@fullsend-ai-coder

Copy link
Copy Markdown
Contributor Author

🔧 Fix agent — iteration 1 (human-triggered)

Applied all 15 items from the ADR 0155 fix round 1 comment to the ADR and docs/architecture.md in one commit with no sign-off. The pre-commit hooks and the link checker could not run in the sandbox because there is no network access; the secret scan passed.

Fixed (7):

  1. Context paragraph 1 replaced (items 1) (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Replaced the first Context paragraph with the plain-words text (no hash, path or Go identifiers), kept the code-review link and mentioned feat(harness): file-level overrides within skill directories #6158 only there; Context's later 'base branch' became 'trusted ref'.
  2. Decision items 1-3 (string shorthand, overrides before discovery and where a new persona goes, source trust rules) (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Added the string shorthand and default-key rule; described overrides as applied before persona discovery and where an added persona lands, with name matching the key; reworded source as a trust-boundary field accepted only from the trusted ref, resolved against the same .fullsend/ directory as skill overrides, with sha256 pin, allowlist and lock behaviour.
  3. Decision items 4-5 (enabled:false guarantee, layering) (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Stated that not staging the file is the guarantee and the dimension-based finding drop is an agent-reported backstop; added ADR 0104's shared-preset rule still holding for model and source entries.
  4. Decision item 8 and new runtimes item, YAML example (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Removed function and file names and stated the three undiscovered-persona cases, with a warning in the run log for enabled:false on an unshipped persona; added item 9 on runtimes; example now uses style-conventions and the raw-content URL form, with comments consistent with item 8.
  5. Forge scope and Consequences (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Replaced forge scope with the requested wording (no claim that Forgejo works today); kept five one-sentence-style bullets, replaced the CLI bullet without flag names, aligned bullet 2 with the backstop wording, and replaced the last bullet.
  6. Living doc: architecture.md bullet (docs/architecture.md): Aligned the finding-drop wording with the backstop framing and changed 'base branch' to 'trusted ref'.
  7. Final grep for hashes, paths and Go identifiers (item 15) (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Grepped Context and Decision; no commit hash, Go source path or Go identifier remains. Product paths required by the instructions (sub-agents/.md, .fullsend/) are kept.

Tests: passed

CI inspections (4):

  1. CI / test (pending, in_progress): Still running when inspected; this change is documentation only.
  2. CI / commit-lint (passing, success)
  3. E2E Tests / playback (passing, success)
  4. Build Site / build (passing, success)

Conflict update: skipped; strategy none; forge MERGEABLE; target main

Next steps:

  • /fs-review — request a re-review of the changes
  • /fs-fix <your instruction> — run another fix pass with specific guidance
  • Push commits directly — review re-runs automatically on push
    Updated by fullsend fix agent

@fullsend-ai-review

fullsend-ai-review Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 4:16 PM UTC · Completed 4:24 PM UTC

Commit: 1a4993a · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $1.46

@waynesun09

Copy link
Copy Markdown
Member

ADR 0155 fix round 2 (concision)

Do not sign off: no git commit -s, no Signed-off-by trailer.

Round 1 made the ADR 21% longer by appending caveats. Replace the second Context paragraph, the whole Decision section (including the YAML block) and the whole Consequences section of docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md with the text below, verbatim. Leave everything else as is: frontmatter, the first and last Context paragraphs, Options and Forge scope. Rewrap lines to the file's usual width. Change no other files.

Context, second paragraph (replace)

The review skill already departs from the scripted pipeline of
[ADR 0018](0018-scripted-pipeline-for-multi-agent-orchestration.md) by relying
on a fixed roster. The roster stays fixed for each run: it is the staged set,
read from the trusted ref, not something the change's author picks. This ADR
does not supersede ADR 0018.

Decision (replace the whole section body)

We extend `agents[].subagents` so that each persona entry is a string or an
object, to let a repository replace, add or disable a persona without forking
the skill.

1. **Two forms.** An entry is a string or an object `{model, source, enabled}`
   with every field optional. A string is shorthand for `{model: <value>}`, so
   ADR 0104's resolution is unchanged and still applies only where it is
   implemented (pi, and codex when the agent delegates). The `default` key
   takes only the string form.
2. **Compiles into skill file overrides.** At staging, before persona
   discovery and on every runtime, the runner turns `source` and
   `enabled: false` into harness skill file overrides. These replace or remove
   `sub-agents/<persona>.md` in the skill that ships the persona. A `source` for
   a persona that no staged skill ships adds the file to the agent's first
   skill that has personas, and the file's frontmatter `name` must equal the
   key.
3. **`source` crosses a trust boundary.** Like the guarded fields of
   [ADR 0112](0112-overlays-may-set-any-harness-field.md), it is read only from
   config at the trusted ref. It is never read from an overlay, a flag, an
   environment variable or the change under review, so a pull request cannot
   change its own review. It is either a path that stays inside `.fullsend/`
   (no `..`, no absolute path, no symlink escape), or a URL pinned with
   `#sha256=`, on `allowed_remote_resources`, with a lock entry that
   `fullsend lock` writes and the run checks.
4. **`enabled: false` is a host guarantee.** The persona file is not staged,
   so no child can be dispatched as that persona. As a backstop, the review
   post-script drops findings whose agent-reported `dimension` names a persona
   that was not staged. This filters mislabelled output; it is not proof of
   what ran.
5. **Layering.** Layers merge per field, and `key: ~` removes an inherited
   entry (ADR 0104). ADR 0104's rule against `subagents` entries in shared
   presets still holds for `model` and `source`.
6. **No persona semantics in the runner.** fullsend checks shape only: types,
   the `source` trust rules and duplicate names. The skill decides what to do
   when, for example, `correctness` is missing.
7. **Selection is unchanged.** The review agent still chooses which dimensions
   run, now only from the staged personas.
8. **Unknown keys.** A key without `enabled: false` that names a persona the
   run did not discover stays an error, unless its `source` defines that
   persona. An entry with `enabled: false` never errors, and its other fields,
   inherited ones included, are ignored. When no skill ships that persona, the
   run log gets a warning. An org-wide disable then works in every
   repository, and a typo shows up.

```yaml
agents:
  - name: review
    subagents:
      default: sonnet              # string form, as in ADR 0104
      style-conventions:
        enabled: false             # not staged
      correctness:
        source: .fullsend/personas/correctness.md
      licensing:                   # added: source defines the persona
        model: opus
        source: https://raw.githubusercontent.com/example-org/personas/<sha>/licensing.md#sha256=<sha256>
```

Consequences (replace the whole section body)

- A repository can add, replace or disable a review persona from config
  without copying the skill, and keeps upstream updates for everything else.
- Disabling is guaranteed on the host because the persona is not staged,
  while choosing dimensions stays the review agent's judgement.
- A `source` change crosses a trust boundary and gets the same review,
  pinning and locking as a skill or harness base.
- The CLI gains a way to set `source` and to disable a persona, implemented
  after this ADR is approved.
- Disabling or adding a persona takes effect only once the review skill builds
  its roster from the staged persona files and fills in `dimension`, which is
  agents-repository work that must ship first.

This round also resolves the review bot's MEDIUM (finding f_7779994ac02e421d: enabled:false plus an inherited model) through item 8, and its LOW through item 4 and the last Consequence.

@waynesun09

Copy link
Copy Markdown
Member

/fs-fix Replace the ADR 0155 sections verbatim with the text in the "ADR 0155 fix round 2" comment: #8969 (comment)
Do not sign off: no git commit -s, no Signed-off-by trailer.

@fullsend-ai-coder

fullsend-ai-coder Bot commented Oct 31, 2026 •

Copy link
Copy Markdown
Contributor Author

🤖 Finished Fix · ✅ Success · Started 4:19 PM UTC · Completed 4:23 PM UTC

Commit: 1a4993a · View workflow run →

Runtime: claude · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $0.29

@fullsend-ai-coder

Copy link
Copy Markdown
Contributor Author

🔧 Fix agent — iteration 2 (human-triggered)

Replaced the ADR 0155 second Context paragraph, Decision section and Consequences section verbatim with the round 2 comment text. No other files changed.

Fixed (1):

  1. Replace ADR 0155 second Context paragraph, Decision and Consequences with the round 2 text (docs/ADRs/0155-agent-spec-subagents-source-or-disable-personas.md): Replaced the three sections verbatim with the supplied concise text; frontmatter, first and last Context paragraphs, Options and Forge scope left unchanged.

Tests: passed

CI inspections (2):

  1. CI / test (passing, in_progress): Still running on the previous head; no failure to diagnose.
  2. CI / commit-lint, web, gateway-doc-configs, test-sandbox-darwin; Build Site / build; E2E Tests; PR Title Lint; DCO (passing, success): All completed checks passed.

Conflict update: skipped; strategy none; forge MERGEABLE; target main

Next steps:

  • /fs-review — request a re-review of the changes
  • /fs-fix <your instruction> — run another fix pass with specific guidance
  • Push commits directly — review re-runs automatically on push
    Updated by fullsend fix agent

@fullsend-ai-review

fullsend-ai-review Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ❌ Failure (fullsend config could not be loaded, agent did not run: loading fullsend config: config.yaml: create_issues is no longer supported: move create_issues.allow_targets to tracker.<provider>.allowed_ta…) · Started 4:10 PM UTC · Completed 4:10 PM UTC

Commit: 614dfd0 · View workflow run →

fullsend-ai-coder Bot and others added 12 commits November 10, 2026 14:20
Add ADR 0155. It records the decision that agents[].subagents entries
may be either today's model string or an object {model, source,
enabled}. At staging, source and enabled: false compile into the skill
file overrides from #6158, so a repository can replace, add or disable
a review persona without forking the orchestrator skill.

The ADR makes source a guarded field (ADR 0112): an in-repo path or a
sha256-pinned URL on allowed_remote_resources, read from the base
branch. A disabled persona is enforced on the host: it is not staged,
and the review post-script drops findings for unstaged dimensions. It
also relaxes the unknown-persona error for enabled: false entries.
ADR 0104 is not edited.

Living documents updated with surgical additions: a Decided bullet in
docs/architecture.md, a contextual link in docs/problems/code-review.md,
and a not-yet-implemented note in docs/reference/config-reference.md.
This is a decision record only; implementation is tracked separately.

pre-commit could not fetch remote hook repos in the sandbox (TLS
error), so hooks ran directly: ADR status/number/frontmatter lints,
docs link lints, lychee, gitleaks and whitespace/EOF checks all pass.

Closes #8968

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rewrite ADR 0155 Context and Decision in plain words (no hashes, file
paths or Go identifiers), tighten the source trust rules, reword the
enabled:false guarantee with the dimension-drop backstop, relax the
undiscovered-persona error per case, add runtimes and forge-scope
clarifications, and align docs/architecture.md.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Replace the second Context paragraph, the Decision section and the
Consequences section with the concise text from the round 2 comment.
Resolves the enabled:false plus inherited model finding via Decision
item 8, and the dimension backstop finding via item 4 and the last
Consequence.

pre-commit could not run (hook repos unreachable: certificate error), and
lychee is unavailable in the sandbox. The ADR status, number, frontmatter
and docs-links hooks were run directly and passed.

Addresses #8969

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Replace Decision items 2, 3, 5 and 6 with the text from the fix round 3
comment, and narrow the architecture.md bullet to a path inside
`.fullsend/`.

pre-commit could not run (hook repos unreachable: certificate error), and
lychee is unavailable in the sandbox. The secret scan passed.

Addresses #8969

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
ADR 0155 fix round 4: note in decision item 2 that the shared lock step
lists pinned skills' sub-agents/ files and partly closes ADR 0104's
deferred lock gap, and in item 3 say the guarded fields are stricter than
ADR 0112's and read from the trusted ref and its pinned preset layers.

Note: pre-commit could not fetch hook repositories (certificate
verification failed in the sandbox); trailing-whitespace, line-ending and
secret checks were run directly. lychee is unavailable in the sandbox, so
make lint-md-links was not run; no links were changed.

Addresses #8969

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Replace the object-form block under subagents in config-reference.md
with a one-line pointer to ADR 0155, since the shape is not yet
accepted by the product.

Pre-commit hooks could not run (network/cert failure fetching hook repos).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Reword ADR 0155 Decision item 7: disabling a persona guarantees its
prompt is not staged, not that no child can be dispatched under its name
(ADR 0104 still allows an anonymous child). Applies the edit requested in
the 'ADR 0155 fix round 7' comment.

Note: pre-commit could not run (hook repos unreachable: TLS verification
failed); checked trailing whitespace, line endings, and EOF newline by hand.

Addresses #8969

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Replace the ADR 0155 bullet in docs/architecture.md verbatim with the
text from the "ADR 0155 fix round 8" comment (finding f_62bf6730c85620d4):
overrides compile at harness composition, trusted-ref rule covers both
fields, and a disabled persona's child runs only as an anonymous helper.

Note: pre-commit could not fetch hook repos (certificate failure) and
lychee is unavailable in the sandbox; verified trailing whitespace, CR
and the ADR link target manually.

Addresses #8969

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Harness skills: file overrides now apply at staging (#6158). Context
and the options list said they did not, and framed the ADR as a way
to avoid forking the skill. Restate the gap: overrides are keyed by
file path, so 0155 adds a per-persona config entry compiled into them.

Assisted-by: Claude
Signed-off-by: Wayne Sun <gsun@redhat.com>
ADR 0162 lets an agents entry carry a harness: block, so subagents
becomes a harness field, set as agents[].harness.subagents or in a
custom harness file. The entry field stays a deprecated alias for
ADR 0104 model strings; setting both in one layer fails.

- Compiled overrides win over skills: after all harness layers merge.
- No overlay can set subagents.*.source or .enabled, block overlays
  included; persona models follow ADR 0112 like model.
- key: ~ keeps removing an inherited entry, as an explicit exception
  to 0162's replace-not-remove rule; enabled: true re-enables.
- The ADR 0104 preset ban applies to the config.base.yaml block.
- source paths resolve like other harness paths; a URL source on a
  built-in agent fails, since lock does not lock built-ins.
- Renamed the file to drop "agent-spec"; depends on ADR 0162.

Assisted-by: Claude (author), Gemini (decision and prose review)
Signed-off-by: Wayne Sun <gsun@redhat.com>
ADR 0162 merged in #8965, so the two links to it now point at the
file instead of the pull request.

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

fullsend-ai-review Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ❌ Failure (post-script /home/runner/work/fullsend/fullsend/.fullsend/.fullsend-cache/resources/sha256/8705f4e8b8e0f68615e6e4514ebf165f4121d0f94d3239c43fcf653157463931/scripts/post-review.sh failed: exit status 1) · Started 7:23 PM UTC · Completed 7:32 PM UTC

Commit: 77873f1 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $1.38

@fullsend-ai-review fullsend-ai-review Bot removed the ready-for-merge All reviewers approved — ready to merge label Nov 10, 2026
@fullsend-ai-review

fullsend-ai-review Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 7:49 PM UTC · Completed 8:00 PM UTC

Commit: 77873f1 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $2.28

fullsend-ai-review[bot]

This comment was marked as outdated.

The "Decided, not yet implemented" callout in harness-fields.md named
ADRs 0112, 0158, 0151 and 0162 but not 0155, which makes subagents a
harness field. Add it in the same style.

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

fullsend-ai-review Bot commented Nov 11, 2026 •

Copy link
Copy Markdown

🤖 Finished Review · ✅ Success · Started 2:06 PM UTC · Completed 2:13 PM UTC

Commit: 4c99d51 · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $1.48

@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.

@@ -0,0 +1,212 @@
---
title: "155. Sub-agent entries may source or disable personas"

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

[medium] pattern-violation

ADR number 0155 is already taken on current main by docs/ADRs/0155-configure-vault-backed-runner-secrets.md ("155. Select native or Vault sources for runner secrets"). Adding a second 0155 record breaks the unique-number check in hack/lint-adr-numbers, so lint will fail once this is merged with main. Any reference to "ADR 0155" would also point to two different decisions.

Suggested fix: Pick an unused ADR number (0154 and 0156 are also taken on main), rename the new file, and update its frontmatter title and H1. Then update every link and ADR-number reference this PR adds in architecture.md, harness-fields.md, problems/code-review.md and config-reference.md. Leave the existing accepted Vault ADR unchanged.

This branch was successfully deployed

1 active and 1 inactive deployments
site-preview — 4c99d51b Deployed Nov 11, 2026 by github-actions[bot]
dev — 4c99d51b Deployed Nov 11, 2026 by waynesun09 via playback #19687
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

component/docs User-facing documentation fullsend-no-fix Skip bot-triggered fix agent runs ready-for-review Triggers review agent dispatch risk/moderate PR risk: moderate type/adr

Projects

None yet

Development

Successfully merging this pull request may close these issues.

docs(adr): let agent-spec sub-agent entries source or disable personas

2 participants