[lukemaj]

What: Structured GitHub CLI calls keep JSON free of terminal color, and dispatch now reports actionable errors for empty or malformed stdout.
Why: Forced color/TTY settings contaminated JSON; dispatch also returned null for empty output or leaked a bare JSONDecodeError for malformed output.
So what: The coordinator owns the next independent review of this exact head and any merge decision; this is an implementation handoff.

Closes #93

Objective contribution: reliable structured GitHub output preserves trustworthy task evidence.

Changes:

• Both structured callers use a copied child environment with NO_COLOR=1, CLICOLOR=0, CLICOLOR_FORCE=0 and GH_FORCE_TTY removed. Parent environment and payload content are preserved.
• Dispatch names the shell-quoted command and stderr (or (no stderr)) when successful gh returns empty, whitespace-only or non-JSON stdout. Malformed JSON retains JSONDecodeError as its explicit cause. Environment tokens are not included in the diagnostic.
• Publish already provides command/stderr diagnostics for empty and malformed output; its existing real-subprocess tests pass. Its fix: explain PR task form for empty Issue publish scopes #110 empty-scope changes remain intact.
• Added dispatch regression coverage for all three invalid-output forms, with and without stderr, token exclusion and exception chaining. Existing synthetic subprocess tests cover forced color, forced TTY and payload preservation.

Candidate:

• Base: current origin/main f57824ea180194c317753f99469e9722213840eb (includes fix: report when trace results are truncated #109-fix(claude): recognize native request-replay metadata #113).
• Head: f82c419969b23e5aecd7231b4b222e4136a058fc on fix/93-gh-structured-color.
• Rebase completed without conflicts. Force-pushed only this branch using an explicit lease against previous head 476adee3feda4bae154729199107279981ece518; remote and PR head verified afterward.

Proof on the exact committed head:

• Failing first: python3 -m unittest -v tests.test_gh_environment.StructuredGhEnvironmentTests.test_dispatch_invalid_stdout_names_command_and_stderr on the rebased pre-fix code produced 2 failures (empty output returned null) and 4 errors (bare JSONDecodeError). The same test passes after the fix.
• python3 -m unittest -v tests.test_gh_environment tests.test_publish_gh tests.test_dispatch tests.test_publish: 41 passed.
• python3 -m unittest discover -s tests: 810 passed in 71.944s, Python 3.14.7.
• python3 -m compileall -q agent_observer tests scripts: passed.
• bin/agent-observer --version: 1.2.6; git diff --check: passed; worktree clean.
• Exact-head CI: success.
• Prior forced-color/TTY failing-first subprocess evidence remains recorded in fix: keep remaining structured gh output free of terminal color #93 for the original implementation.

Luna r1 full-suite discrepancy:

• Full suite on current main f57824ea180194c317753f99469e9722213840eb, in an isolated checkout with Git history: 806 passed in 63.568s, same host/Python. None of the four reported failure areas reproduced on main or the candidate.
• tests.test_placement, tests.test_tool_arguments, tests.test_writer_lock: all 40 tests also passed in 4.777s with BOTH dispatch/publish structured_env entry points patched to raise if invoked. No invocation occurred.
• Those implementation and test files are unchanged from main. structured_env copies os.environ and is used only for gh children; placement git and writer-lock subprocess paths do not use it.
• Load sensitivity is plausible: the argument timing assertion is 1s, placement Git subprocess timeout is 5s, and lock tests use bounded 2s/5s waits. Host load averages during investigation were approximately 5-8 on 8 CPUs. The earlier failures' exact environmental cause remains unconfirmed; no evidence supports a regression from this PR, so no unrelated fixes were made.
• An initial baseline archive run had nine historical-upgrade setup errors because it lacked Git history. It was replaced by the successful history-backed baseline above; the archive run is not claimed as full baseline proof.

Limits:

• No independent review was run by this implementation sub-agent. Luna r1's P2 is addressed; the new exact head still needs coordinator-owned independent review.
• No merge, VERSION, plugin manifest, CHANGELOG, workflow/settings, installation, live-ledger writes or publication POSTs. Versioning remains deferred per explicit task instructions.
• Cost comment remains unverified: the earlier isolated-ledger attempt reported nothing to publish: no sessions in scope. No new publish attempt was made.
• Website impact: none, this changes CLI child environments and diagnostics only.

Requirements and who asked: User, #93 and Luna r1 require plain structured output, actionable command/stderr errors, failing-first proof, parent/content preservation, current-main rebase and investigation of four reported suite failures.
Deleted: ANSI stripping, shared parser refactoring, unrelated flaky-test fixes without causal evidence, live publication proof and release-file changes.
Bottleneck: Invalid gh stdout lacked actionable dispatch diagnostics; exact-head and baseline proof now distinguish that defect from the unconfirmed suite discrepancy.
Checked myself: Read #93, #115 and Luna r1; inspected both structured callers and full diff; observed failing-first diagnostics, 41 focused passes, 810 candidate passes, 806 main passes, 40 helper-exclusion passes, compileall, CI and remote identity.

Implemented with Codex (gpt-6.1-sol, medium reasoning) through T3 Code.

[lukemaj]

Luna review r1 of 476adee: REQUEST CHANGES

• P2 — agent_observer/dispatch.py:259: If gh exits successfully but emits malformed or empty stdout (for example, a future CLI/API response or wrapper writes a diagnostic to stdout), _gh_json lets json.loads raise a bare JSONDecodeError. Unlike publish._gh, it omits the command and stderr, so the failure is not actionable. Wrap the parse error with the command and stderr while preserving the cause/context.

The environment helper itself is sound: it copies the parent environment (retaining PATH, GH_TOKEN, GH_HOST, and proxy variables), overrides terminal color flags and removes GH_FORCE_TTY only for the child. Both structured callers use it; the PR leaves VERSION, manifests and CHANGELOG untouched. The subprocess tests cover both forcing cases and verify parent environment/content preservation.

[lukemaj]

Luna review r2 of f82c419: APPROVE

No actionable findings.

• The P2 in dispatch._gh_json is fixed: empty/whitespace stdout raises an actionable RuntimeError; malformed stdout includes the command and stderr and preserves JSONDecodeError as the cause.
• Regression proof: tests.test_gh_environment.StructuredGhEnvironmentTests.test_dispatch_invalid_stdout_names_command_and_stderr passes on this head and produced 2 failures and 4 errors when run against the pre-fix behavior. The error assertions include a synthetic GH_TOKEN and confirm it is absent from the message.
• The previously failing areas pass individually: placement (19), tool arguments (14), writer lock (7). The complete suite passes (810 tests).
• The publish.py refactor retains the behavior merged in fix: explain PR task form for empty Issue publish scopes #110. No VERSION, manifest, or CHANGELOG changes.