chore(#1193): repin harness images to fullsend 0.42.0 digests

[fullsend-ai-coder]

Summary

• Repin all seven harness/*.yaml image digests from the previous release to the fullsend 0.42.0 release
• fullsend-code (code, fix, review) → sha256:623fc74588876062f12f40690ee759c6134ba516fabdd36e6dcfc4eeb9cd1d0b
• fullsend-sandbox (prioritize, retro, scribe, triage) → sha256:259605fea321353552fdefd3a6a55e8b5c260998dfc5a622ed143e41a429995a

Digest table

Image	Digest	Files
ghcr.io/fullsend-ai/fullsend-code:0.42.0	sha256:623fc74588876062f12f40690ee759c6134ba516fabdd36e6dcfc4eeb9cd1d0b	harness/code.yaml, harness/fix.yaml, harness/review.yaml
ghcr.io/fullsend-ai/fullsend-sandbox:0.42.0	sha256:259605fea321353552fdefd3a6a55e8b5c260998dfc5a622ed143e41a429995a	harness/prioritize.yaml, harness/retro.yaml, harness/scribe.yaml, harness/triage.yaml

Both digests resolved from GHCR (Docker-Content-Digest for the 0.42.0 tag); both images carry org.opencontainers.image.revision = 4665685a1d9d471a7a044908a68f00bed78caba7.

What the new image ships

• pi 0.85.0, xai-vertex 0.2.1, Claude Code 2.1.260
• fullsend anthropic-vertex extension
• Runner-side persona routing (feat(runtime): per-persona model configuration for pi sub-agents (#7031) fullsend#7038)
• Hook fix (fix(#7043): keep analysis_incomplete blocking and fail closed on malformed tirith output fullsend#7044)

Validation

$ grep -h '^image:' harness/*.yaml | sort -u
image: ghcr.io/fullsend-ai/fullsend-code@sha256:623fc74588876062f12f40690ee759c6134ba516fabdd36e6dcfc4eeb9cd1d0b
image: ghcr.io/fullsend-ai/fullsend-sandbox@sha256:259605fea321353552fdefd3a6a55e8b5c260998dfc5a622ed143e41a429995a

Related

• release: open the agents harness image-digest repin PR from the tag build instead of by hand (and tag agents after it lands) fullsend#7048 (automate this)
• chore(harness): repin sandbox and code images to the 0.40.0 release digests #1160 (previous repin for 0.40.0)

---

Closes #1193

Post-script verification

• Branch is not main/master (agent/1193-repin-images-042)
• Secret scan passed (gitleaks — ec2f8a5b4369b844323ee1ee87c1d4a599a42817..HEAD)
• PR body secret scan passed (gitleaks — no-git)

[waynesun09]

/fs-fix-stop

[github-actions]

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

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 8:54 PM UTC · Completed 9:06 PM UTC

Commit: 7bb0eab · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $2.60

[fullsend-ai-review]

Risk Assessment: moderate (2/5)

Details

A surgical, bot-authored config-only PR repinning image digests in 7 harness YAML files per a well-scoped issue; the mechanical nature and reversibility of the change offset the high churn, multi-author contention, and regression history observed in these frequently-touched files.

[fullsend-ai-review]

Review

Findings

Medium

• [protected-path] harness/ — This PR modifies 7 files under protected paths: harness/code.yaml, harness/fix.yaml, harness/prioritize.yaml, harness/retro.yaml, harness/review.yaml, harness/scribe.yaml, harness/triage.yaml. The PR links to issue harness: repin sandbox and code images to the 0.42.0 release digests #1193 and explains the rationale (repinning image digests to fullsend 0.42.0). Human approval is always required for protected-path changes, regardless of context.

[waynesun09]

Pin verified independently against GHCR: fullsend-code:0.42.0 -> sha256:623fc74588876062f12f40690ee759c6134ba516fabdd36e6dcfc4eeb9cd1d0b, fullsend-sandbox:0.42.0 -> sha256:259605fea321353552fdefd3a6a55e8b5c260998dfc5a622ed143e41a429995a. All 7 harness files (code, fix, prioritize, retro, review, scribe, triage) carry exactly those two digests; the diff touches only the image: line in each. All checks green.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 7:22 PM UTC · Completed 7:30 PM UTC

Commit: 7bb0eab · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $2.23

[fullsend-ai-retro]

Retro: PR #1194 — repin harness images to fullsend 0.42.0 digests

Timeline

1. Issue #1193 created by waynesun09 (Sep 5, 20:40 UTC) — the seven harness/*.yaml files still pinned previous-release digests after the 0.42.0 release.
2. Triage agent finished in ~1 min ($0.51) — correctly assessed severity as high.
3. Code agent finished in ~6 min ($1.61) — produced PR #1194 with correct single-line image: updates across all 7 harness files.
4. Human posted /fs-fix-stop within 30 seconds of PR creation, preemptively disabling the fix agent.
5. Review agent ran for ~13 min ($2.60, high effort, claude-opus-4-6) — classified the change as mechanical, assessed risk as moderate (2/5), and emitted one medium finding: harness/ is a protected path requiring human approval. Verdict: comment-only.
6. Human approved 3 days later (Sep 8 — weekend gap) with thorough independent GHCR digest verification.
7. PR merged by waynesun09 on Sep 8.

Total agent cost: $4.72. Zero rework iterations. Code was correct on first attempt.

Assessment

This workflow executed well. The code agent produced a clean, correct PR. The review agent's protected-path governance finding was appropriate — harness/ changes should require human sign-off. The 3-day resolution time reflects a weekend gap, not agent inefficiency.

The main optimization opportunities — review cost on mechanical changes, automating the repin process, and skipping retro on trivial PRs — are all tracked by existing open issues. No new proposals are warranted.

Evidence for existing issues

• agents#301, agents#513 (lightweight review path for mechanical/dependency-bump PRs): This PR is a concrete case where the review agent spent $2.60 and 13 minutes (including a challenger pass) to review 7 single-line digest swaps it classified as "mechanical." A lightweight path would have saved ~$2 and ~10 minutes.
• agents#741 (assess semantic significance of protected-path changes before assigning high severity): The PROTECTED_PATH_COUNT scored 5/5 (maximum) because ≥2 files triggered the threshold, despite all 7 files being in the same directory with identical single-line changes. Homogeneity of changes isn't factored into severity.
• fullsend#6607 (automate the agents harness image digest repin after each release): This is the second manual repin in recent history (after #1160 for 0.40.0), further confirming the automation need.
• agents#649 (skip retro on trivially-handled mechanical PRs): This retro itself is low-value — the workflow executed correctly with zero rework on a purely mechanical change.
• fullsend#6933 (cache write failure in per-repo dispatch): The review workflow log shows cache write denied: token has no writable scopes.
• fullsend#6670 (stop-fix job token lacks collaborator API permission): The code workflow log shows collaborator permission lookup failed for fullsend-ai-coder[bot].