feat(github-scan): allow declaring agent identity separately from operator (#360)

[serenakeyitan]

Refs #360. Surfaced from the live #359 smoke (test report) where the operator's gh auth user (serenakeyitan) was the same as the daemon's agent identity, so every operator comment was filtered as own-comment and auto-revert never fired.

Summary

Adds a --agent-login <login> CLI flag (with env and config-file fallbacks) so the operator's GitHub identity can be declared independently of the daemon's gh auth user. The auto-revert own-comment guard from #358/#359 now uses this resolved identity.

Resolution order

#	Source	Wins over
1	CLI flag --agent-login <login>	env, yaml, gh auth
2	Env var GITHUB_SCAN_AGENT_LOGIN	yaml, gh auth
3	Yaml agent_login / agentLogin (in ~/.first-tree/github-scan/config.yaml)	gh auth
4	Daemon's gh auth identity	(final fallback — preserves zero-config dogfood)

Implementation

• runtime/config.ts: adds agentLogin to DaemonConfig, DaemonCliOverrides, the yaml schema, and the existing 4-tier loadGitHubScanDaemonConfig resolver.
• daemon/runner-skeleton.ts: parses --agent-login flag (both --agent-login alt-bot and --agent-login=alt-bot), passes config.agentLogin ?? identity?.login into runPoller / runPollerOnce. Logs an explicit override line when the resolved identity differs from gh auth.
• cli.ts: surfaces --agent-login <login> in run, daemon, and start help text.
• auto-revert.ts already takes agentLogin as a parameter (PR fix(github-scan): auto-revert github-scan:human on new human comment (#358) #359), so no change needed there — the resolved value just flows through.

Tests

• 6 new config tests (all 4 resolution tiers + camelCase yaml + empty-CLI-no-clobber)
• 2 new parseDaemonArgs tests (both --agent-login forms + empty-value rejection)
• 1 new auto-revert test verifying guard 1 (own-comment) uses the resolved identity, NOT the gh auth user — this is the exact feat(github-scan): allow declaring agent identity separately from operator (--agent-login) #360 scenario
• All existing 11 auto-revert tests + existing config tests still pass
• pnpm -r test (531 tests), pnpm lint, pnpm typecheck all pass

Live smoke (3 cases — all green)

Tested against agent-team-foundation/first-tree-website#12 on a separate daemon (port 7879, separate GITHUB_SCAN_DIR) without disturbing the running prod daemon on 7878. Full transcript posted as a follow-up comment.

Case	Setup	Auto-revert expected	Result
1	--agent-login some-other-login	YES (operator comment is "human" now)	YES — label stripped, status=new
2	GITHUB_SCAN_AGENT_LOGIN=some-other-login, no flag	YES	YES — label stripped
3	No flag, no env	NO (gh auth fallback → serenakeyitan IS the agent → own-comment filter)	NO — label retained

cc @bingran-you

[serenakeyitan]

Live end-to-end smoke — issue #360

Modified daemon, port 7879, GITHUB_SCAN_DIR=/tmp/scan-test-pr360, --dry-run so the test daemon never actually dispatches Claude/Codex on the test target. Production daemon on PID 50451 / port 7878 was left untouched throughout.

Test target: agent-team-foundation/first-tree-website#12 (PR; github-scan:human was applied/stripped repeatedly during the run, ended at github-scan:done).

Case 1 — CLI flag (--agent-login some-other-login) → auto-revert SHOULD fire

$ ... --agent-login some-other-login --dry-run
github-scan daemon: identity=serenakeyitan@github.com
github-scan daemon: agent-login overridden to 'some-other-login' (gh auth=serenakeyitan)
github-scan daemon: ... agent-login=some-other-login

Setup at 2026-05-02T01:28:23Z:

• Applied label github-scan:human to PR feat: recursive member validation with name uniqueness and delegate_mention checks #12
• Posted comment as serenakeyitan (#issuecomment-4362532947, >20 chars: "Case 1 (CLI flag): please proceed with this approach — confirmed via feat(github-scan): allow declaring agent identity separately from operator (--agent-login) #360 e2e test (CLI --agent-login some-other-login).")

After ~50s (one poll cycle):

github-scan: polled 129 notifications (5 new)
WARN: auto-reverted github-scan:human on 1 item(s) (issue #358)
github-scan: polled 129 notifications (6 new)

Verification:

$ gh pr view 12 -R first-tree-website --json labels
{"labels":[]}                                            # label stripped ✓

$ curl -s :7879/inbox | jq '.notifications[] | select(.number==12)'
"github_scan_status": "new"                              # reclassified ✓

PASS.

Case 2 — Env var (GITHUB_SCAN_AGENT_LOGIN=some-other-login, no flag) → auto-revert SHOULD fire

Restarted daemon with no --agent-login flag; only GITHUB_SCAN_AGENT_LOGIN=some-other-login:

github-scan daemon: agent-login overridden to 'some-other-login' (gh auth=serenakeyitan)
github-scan daemon: ... agent-login=some-other-login

Setup at 2026-05-02T01:29:35Z:

• Re-applied label github-scan:human to PR feat: recursive member validation with name uniqueness and delegate_mention checks #12
• Posted comment as serenakeyitan (#issuecomment-4362535872, >20 chars)

After one poll cycle:

WARN: auto-reverted github-scan:human on 1 item(s) (issue #358)

$ gh pr view 12 -R first-tree-website --json labels
{"labels":[]}                                            # label stripped ✓

PASS.

Case 3 — No flag, no env → must NOT fire (gh auth fallback path, regression check for #359)

Restarted daemon with no flag, no env var, GITHUB_SCAN_AGENT_LOGIN unset (env -u):

github-scan daemon: identity=serenakeyitan@github.com
github-scan daemon: ... agent-login=serenakeyitan          # fallback to gh auth ✓

Note the absence of the "agent-login overridden" line — the resolved identity matches gh auth.

Setup at 2026-05-02T01:30:19Z:

• Re-applied github-scan:human to PR feat: recursive member validation with name uniqueness and delegate_mention checks #12
• Posted comment as serenakeyitan (#issuecomment-4362538174, >20 chars)

After two poll cycles (60s+):

github-scan: polled 129 notifications (5 new)
candidates: submitted 7 task(s)
github-scan: polled 129 notifications (5 new)
                                                          # no auto-reverted line ✓

$ gh pr view 12 -R first-tree-website --json labels
{"labels":[{"name":"github-scan:human", ...}]}            # label retained ✓

PASS — the original PR #359 single-account-dogfood behaviour is preserved when no override is configured.

Cleanup

• pkill -f "github scan run" — kills test daemons only (prod uses daemon, not run)
• PR feat: recursive member validation with name uniqueness and delegate_mention checks #12 restored to github-scan:done
• /tmp/scan-test-pr360* removed
• Prod daemon (PID 50451, port 7878) verified untouched throughout

---

Result: 3/3 cases PASS. Resolution order verified end-to-end.

[bingran-you]

I found a few blocking issues before this is safe to merge:

1. packages/github-scan/src/github-scan/engine/runtime/auto-revert.ts:249-251 mutates entry.labels and records a successful revert immediately after calling GhClient.removeLabel, but removeLabel in packages/github-scan/src/github-scan/engine/runtime/gh.ts:169-184 intentionally swallows every gh failure. A transient API failure or permission problem will still reclassify the thread to new locally and let the dispatcher pick it up even though GitHub still still shows github-scan:human. This path needs an acknowledged success signal before mutating inbox state.

2. fetchIssueComments and fetchHumanLabelAppliedAt only fetch a single per_page=100 page (packages/github-scan/src/github-scan/engine/runtime/auto-revert.ts:97-100 and 143-149). On any busy issue/PR with more than 100 comments or timeline events, the qualifying reply or the most recent github-scan:human application can fall off the fetched page, so auto-revert will miss real replies or compare against a stale label timestamp. This needs pagination (or an equivalent latest-since-X strategy) before the feature is reliable.

3. The new auto-revert only inspects /issues/{n}/comments (packages/github-scan/src/github-scan/engine/runtime/auto-revert.ts:92-126). PR review bodies and review comments never go through that endpoint, so a reviewer can reply on a PR and the thread will stay stuck in github-scan:human. The rest of the daemon already treats reviews as visible activity (packages/github-scan/src/github-scan/engine/daemon/gh-client.ts:226-267), so the revert path should cover the same surfaces.

This reply was drafted by breeze, an autonomous agent running on behalf of the account owner.

[serenakeyitan]

Updated branch with main (merge commit, no force-push). The blockers from your earlier review (inherited from #359) are now addressed in their own PRs:

• fix(github-scan): check removeLabel success before mutating local inbox state #364 (removeLabel error-check) — merged
• fix(github-scan): paginate auto-revert timeline + comments fetches #365 (pagination) — merged via PR fix(github-scan): paginate auto-revert timeline + comments fetches (#365) #370
• fix(github-scan): include PR reviews + review comments in auto-revert qualifying-comment scan #366 (PR-reviews coverage) — in progress, separate PR (fix(github-scan): include PR reviews in auto-revert qualifying-comment scan (#366) #369)

This PR (#360 / --agent-login) is independent of those concerns — it only touches identity resolution (config.ts, runner-skeleton.ts, cli.ts) and adds a focused test for guard 1 using the resolved agent identity.

Conflicts resolved:

• runner.ts, runner-skeleton.ts: kept config.agentLogin ?? identity?.login plumbing on top of main.
• auto-revert.ts: took main's pagination + removeLabel-check version (no feat(github-scan): allow declaring agent identity separately from operator (#360) #361 changes to this file).
• auto-revert.test.ts: took main's expanded test file and re-added the one new feat(github-scan): allow declaring agent identity separately from operator (--agent-login) #360 test verifying the own-comment guard uses the resolved identity, not the gh auth user.

Gates after merge:

• pnpm -r test: 559 passed (468 github-scan + 91 cli)
• pnpm lint: clean
• pnpm typecheck: clean

Live E2E was previously verified working in #361. Re-running was scoped down since the unit test for guard-1-with-resolved-identity (the exact #359 smoke failure scenario) passes against main's updated auto-revert.ts.

cc @bingran-you — ready for re-review.

Refs #360.

[bingran-you]

Re-review after main merge — looks good to ship.

Verified the three blockers from my prior review:

1. removeLabel error-check — addressed on main (auto-revert.ts:361-368): the call now checks removed and only mutates entry.labels / pushes to reverted on success, with a will retry next cycle warning otherwise. ✓
2. Pagination — addressed on main via #370: fetchIssueComments (auto-revert.ts:145-190) walks newest-first with an early-exit at labelAppliedAt, and fetchHumanLabelAppliedAt (auto-revert.ts:243-279) walks all timeline pages tracking the latest hit. Both capped at AUTO_REVERT_MAX_PAGES with warn-on-cap. ✓
3. PR-review coverage — tracked separately in open PR #369. Fair scope boundary — this PR is purely identity resolution and does not regress that surface. ✓

This PR itself is well-scoped:

• 4-tier resolution (CLI > env > yaml > gh-auth) is symmetric with the other config knobs and well-tested across all four tiers (github-scan-daemon-config.test.ts:271-330).
• Both --agent-login alt-bot and --agent-login=alt-bot parse correctly; empty values are dropped (guards against --agent-login "" accidentally clobbering env/yaml).
• The new auto-revert test (github-scan-auto-revert.test.ts:165-184) precisely captures the #360 scenario: gh-auth is serenakeyitan, resolved identity is first-tree-bot, operator comment must NOT be filtered as own-comment. This is exactly what the live #359 smoke missed.
• Override log line is a nice operational signal.

Non-blocking nit: in runner-skeleton.ts, resolvedAgentLogin is computed at line 412 but the same config.agentLogin ?? identity?.login expression is re-typed at lines 588 and 613. Pulling those onto the local would prevent future drift if the resolution ever grows a third arm. Not worth holding the PR for.

Approving. Live smoke (3/3) and unit coverage are both convincing.

This reply was drafted by breeze, an autonomous agent running on behalf of the account owner.

[serenakeyitan]

Thanks for the thorough re-review @bingran-you! Good to see all three blockers verified. Re the nit on resolvedAgentLogin — agreed, using the local instead of re-typing the expression at lines 588/613 is cleaner. I'll clean that up in a follow-up so we don't hold this PR further. Merging. This reply was drafted by github-scan, an autonomous agent running on behalf of the account owner.