Skip to content

chore: roll back to the default claude harness; demote claude-interactive - #698

Merged
max-sixty merged 4 commits into
mainfrom
dash-p
Jun 17, 2026
Merged

max-sixty merged 4 commits into
mainfrom
dash-p

Conversation

@max-sixty

Copy link
Copy Markdown
Owner

Anthropic paused the 2026-06-15 Agent SDK credit metering (the -p billing change), so subscription claude-code-action runs again draw from the plan's usage limits. That removes the billing motivation for the claude-interactive harness.

This reverts tend's own dogfooding to the default claude harness, demotes claude-interactive to a footnote in the docs, and scrubs the now-stale billing caveat.

Changes

  • .config/tend.yaml: drop harness: claude-interactive (the default is claude).
  • Regenerate tend-*.yaml via uvx tend@latest init: the action ref goes from max-sixty/tend/interactive@0.1.4 to max-sixty/tend@0.1.4 across all eight workflows.
  • README, docs/tend.example.yaml (and its synced install-tend mirror), and the install-tend skill: claude-interactive is now a single [^interactive] footnote; the secrets table and harness list show claude + codex only.
  • README and install-tend skill: remove the metered Agent SDK credit caveat, which Anthropic's help center now lists as paused.

The claude-interactive harness still ships and works; it is just no longer the recommended path. The composite action, generator support, and session-log handling are untouched.

This was written by Claude Code on behalf of max

…tive

Anthropic paused the 2026-06-15 Agent SDK credit metering (the `-p` billing
change), so subscription claude-code-action runs again draw from the plan's
usage limits. That removes the billing motivation for the claude-interactive
harness, so revert tend's own dogfooding to the default claude harness and
demote claude-interactive to a footnote in the docs.

- .config/tend.yaml: drop `harness: claude-interactive` (default is claude)
- regenerate tend-*.yaml via `uvx tend@latest init`: max-sixty/tend@0.1.4
- README, docs/tend.example.yaml (+ synced mirror), install-tend skill:
  claude-interactive is now a footnote; secrets table and harness list show
  claude + codex only
- README, install-tend skill: remove the now-stale metered Agent SDK credit
  caveat

The claude-interactive harness still exists and works; it is just no longer
the recommended path.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

@tend-agent tend-agent left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

One leftover billing-caveat reference outside the diff: interactive/action.yaml's description still cites the now-paused June 2026 metering split as the harness's motivation. Suggested replacement for lines 7–9:

  is terminated cleanly. Inputs

That collapses the motivation sentence and lets the folded scalar reflow into the existing "Inputs overlap with action.yaml's …" sentence. Happy to push the fix if you want.

max-sixty and others added 3 commits June 17, 2026 10:30
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The June 2026 Agent SDK billing split that motivated the interactive
harness was paused, so the description no longer cites it. Flagged by the
tend reviewer on #698.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@max-sixty
max-sixty merged commit eabebc7 into main Jun 17, 2026
7 checks passed
@max-sixty
max-sixty deleted the dash-p branch June 17, 2026 17:58
max-sixty added a commit that referenced this pull request Jun 18, 2026
…704)

## What this does

Replaces the default `claude` harness (root `action.yaml`) with a
self-contained headless `claude -p` run, executed as a non-sudo
`tend-sandbox` user behind the same credential-injecting mitmproxy the
`claude-interactive` harness already uses. The bot PAT and the Anthropic
credential live only in the proxy and never enter the agent's
environment or disk. Previously this harness wrapped
`anthropics/claude-code-action@v1`, which hands both secrets to the
agent directly.

The motivation is credential isolation, not billing. (Anthropic paused
the 2026-06-15 Agent SDK metering in #698, so the two Claude harnesses
now bill the same way; they differ only in completion supervision —
headless `-p` exit code vs the interactive harness's PTY Stop-hook
sentinel.)

## How to review

The weight is `action.yaml` (a near-total rewrite) and the new smoke
workflow. The proxy machinery under `interactive/proxy/` is reused in
place, unchanged.

- **`action.yaml`** — the new step sequence: base-branch restore of
RCE/injection-relevant config (`.claude`, `.mcp.json`, `CLAUDE.md`,
`.gitmodules`, …) → sandbox + proxy setup → toolchain install into the
sandbox home → non-interactive plugin install → a supervised `claude -p
--output-format stream-json` run whose completion is the process exit
code plus the final `result` event → workspace ownership handback →
stream-json token parsing. Preflight, notification, and artifact steps
stay in lockstep with `interactive/action.yaml` for swap-in parity.
- **`.github/workflows/claude-smoke.yaml`** — installs and runs the
harness as the sandbox user and deterministically gates on
Anthropic-credential isolation (greps the stream-json, hard-fails on a
reported leak or a missing verdict). It runs on PRs that touch the
harness (`action.yaml` or the shared proxy) and on manual dispatch; the
prompt lives in `.github/claude-smoke-prompt.md`.
- **docs** — `docs/security-model.md`, `CLAUDE.md`, `TODO.md`, README,
the install-tend skill, and `tend.example.yaml` now describe both Claude
harnesses as proxy-isolated and drop the claude-code-action / billing
framing.

## External interface (`action.yaml`)

- Removed inputs: `allowed_bots`, `allowed_non_write_users`,
`use_sticky_comment`, `additional_permissions` (claude-code-action
passthroughs the generator never emitted; engagement and permissions
live in the generated workflow).
- Added inputs: `claude_version` (default `2.1.179`, in step with the
interactive harness), `timeout_seconds` (`3600`), `mitmproxy_version`
(`12.2.1`).
- `token_usage` output shape is preserved; session logs move to
`${RUNNER_TEMP}/tend-logs/`.

## Testing

- Generator suite (255) and the proxy credential-injection suite (21)
pass; lints clean.
- `claude-smoke` ran green on this PR: `anthropic-cred: isolated` (no
real Anthropic secret reachable by the sandboxed agent), `claude -p`
completed 7 turns on sonnet, GitHub PAT and git both reached origin
through the proxy. This is the end-to-end check that the harness has no
unit coverage of.
- The generated `tend-*.yaml` workflows pin `max-sixty/tend@0.1.5`, so
this rewrite does not affect tend's live dogfooding until a release tags
the new action.

> _This was written by Claude Code on behalf of max_

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants