suzuke · suzuke · Apr 27, 2026 · Apr 27, 2026
diff --git a/docs/ts-team-backlog.md b/docs/ts-team-backlog.md
@@ -0,0 +1,17 @@
+# ts-team backlog
+
+Tracking non-blocking gaps surfaced during deprecated-mode maintenance. Items here did not block the sprint they were found in. They are kept for future-sprint pickup, agend-terminal port, or eventual `wontfix` closure.
+
+## Open
+
+### B-001 — Session-routing edge case in cost-guard gate
+
+- **Surface**: `src/outbound-handlers.ts.sendToInstance()` cost-guard pre-check
+- **Surfaced by**: ts-reviewer, PR #57 review (Out-of-scope Observation 2)
+- **Surfaced at**: 2026-04-27, Sprint 0
+- **Priority**: low
+- **Description**:
+  `isLimited(targetName)` is called *before* session resolution at `src/outbound-handlers.ts:101-105`. When `targetName` is an external session name (resolved later via `ctx.sessionRegistry.get(targetName)` at line 115), `CostGuard` returns 0 for the unresolved name and the gate is bypassed — even when the *host* instance is over its daily limit.
+- **Impact**: Cost-guard protection is incomplete on the session-routing path. Probably rare in practice (external user sessions are typically the destination, not over-budget targets), but the gate's protection is non-uniform.
+- **Estimated fix**: ~5 LOC src + 1 test. Resolve to host instance name first, then check `isLimited` against the resolved name.
+- **Disposition**: deferred. Decision boundary: if hit rate stays low through agend-terminal migration window → close as `wontfix-deprecated`. If a real incident lands → reopen and patch.
diff --git a/docs/ts-team-sprint-0-2026-04-27-wrap.md b/docs/ts-team-sprint-0-2026-04-27-wrap.md
@@ -0,0 +1,80 @@
+# ts-team Sprint 0 wrap — 2026-04-27
+
+**Window**: 2026-04-27 01:18 UTC → 03:28 UTC (~2h10m)
+**Repo**: AgEnD (TS), maintenance mode post-deprecation (commit `ca1b729`)
+**Team**: ts-lead, ts-impl, ts-reviewer (3 instances, claude backend)
+**Authority chain**: operator TUI direct dispatch (01:18) → operator overnight delegation telegram (01:52, decision `d-20260427025246010573-0`, 24h TTL) → general acting as strategic surrogate from 01:52 onward
+
+## Mission
+
+Triage AgEnD repo open issues. Bug → discuss + fix. Feature → estimate + report. Per-issue disposition decided in collaboration with general (operator surrogate) once the operator entered the overnight sleep window.
+
+## Outcome
+
+| Issue | Type | Disposition | Evidence |
+|---|---|---|---|
+| #55 | Bug — Gemini fleet context dual-injection | **Fixed** (scope expanded to all dual-injection backends) | PR #56 squash-merged `f58ad9d3` |
+| #24 | Feature — usage-limit notify sender | **Fixed** | PR #57 squash-merged `464c8f40` |
+| (#24 follow-up) | broadcast surface gate gap | **Fixed** (surfaced by PR #57 reviewer) | PR #58 squash-merged `0ec12443` |
+| #52 | Feature — server-side `pause_fleet` primitive | **Closed**, migrate-to-agend-terminal | comment 4323658306 |
+| #53 | Feature — Telegram `/pause` `/resume` slash commands | **Closed**, migrate-to-agend-terminal | comment 4323658307 |
+| #54 | Feature — Telegram multipart envelope | **Closed**, migrate-to-agend-terminal | comment 4323658308 |
+| #8 | Feature — default topic package for new users | **Closed**, migrate-to-agend-terminal | comment 4323658309 |
+
+Net: 0 open issues, 0 open PRs, all six triaged issues resolved (3 fixed + 4 closed-with-migration).
+
+## Bug #55 — fix detail
+
+ts-impl audit during the triage discussion expanded the original Gemini-only issue to a systematic fix:
+
+- Claude backend was *not* in scope (writes `fleet-instructions.md` to instance dir, not workspace, so no project-doc dual-injection — already used `--append-system-prompt-file`).
+- Gemini (`GEMINI.md`), Codex (`AGENTS.md`), and Kiro (`.kiro/steering/agend-<name>.md`) all had the dual-injection pattern.
+- A second silent dual-injection source (`AGEND_DECISIONS` via mcpEnv vs. `buildFleetInstructions({ decisions })`) was caught and folded into the same gate.
+
+Fix shipped Option B generalization (over Option A "MCP-only") — the original issue presented a binary choice; we chose the human-debuggable + non-MCP-dependent path.
+
+- `CliBackend.nativeInstructionsMechanism: 'append-flag' | 'project-doc' | 'none'`
+- `daemon.buildBackendConfig()` drops fleet-context env vars when the backend declares its own native injection
+- MCP server gates the `instructions` capability via `AGEND_DISABLE_MCP_INSTRUCTIONS` env
+
+Net: 12 files / +265 / -11. Tests: 12 new (parametrised across all 6 backends), 517/517 pass post-fix.
+
+## Feature #24 — fix detail
+
+Two-PR sequence:
+
+1. **PR #57** (~30 min, src +23 / tests +184): cost-guard pre-check in `sendToInstance`/`delegateTask`. ts-impl correctly stayed in scope per dispatch.
+2. **PR #58** (~25 min, src +14 / tests +133): broadcast surface, surfaced by ts-reviewer's #57 review as Out-of-scope Observation 1. Per-target gate inside the existing fan-out loop, with new `cost_limited` field in the result envelope (backwards-compat: existing consumers of `sent_to`/`failed`/`count` unaffected).
+
+Net combined: src +37, tests +317, 524→527 tests pass.
+
+Pattern note: dispatching the broadcast follow-up as a separate PR (rather than expanding #57 mid-review) preserved a clean review record and let the original PR ship without churn — same pattern dev-lead has been using on Sprint 19+.
+
+## Closed issues — migration path
+
+All four close comments are individually personalised:
+
+- Each cites a concrete detail from the issue body (incident date for #52/#54, the issue's own dependency declaration for #53, token-discipline angle for #8).
+- Each links forward to `agend-terminal`'s issue tracker.
+- Each acknowledges the value of the proposal — close-with-respect, not close-as-rejected.
+- Wording reviewed by general before posting (`watered-down implementation` → `incomplete implementation` on #53 for technical neutrality).
+
+## Process notes
+
+- **Operator delegation handoff**: operator TUI 01:18 → general overnight 01:52. ts-lead initially declined a Sprint 0 dispatch from general (correctly — no durable instruction yet); general posted `d-20260427025246010573-0` as evidence; ts-lead verified and authority chain accepted. Both sides recorded incident lessons in the decision body.
+- **Worktree discipline**: every PR opened from a fresh worktree (`fix/55-fleet-context-dual-injection`, `feat/24-usage-limit-notify`, `fix/broadcast-cost-guard-gap`). Main never modified directly.
+- **Reviewer Contract v1.1**: every review reported `scope_source`, `audit_mode`, `reviewed_head`, exact commands run, files audited, with explicit out-of-scope observations surfaced separately from blocking findings.
+- **Operator hard rule "務必確認可正確運行"**: applied especially to #24 — three-state mock test (limited / not-limited / null-costGuard) was the test scaffold for both #57 and #58.
+
+## Backlog spawned this sprint
+
+- `docs/ts-team-backlog.md` B-001: session-routing edge case in cost-guard gate (low priority, deferred).
+
+## Next
+
+Sprint 1 — pure-cleanup (no new code on production paths):
+
+- **Track 1**: docs migration consistency audit + deprecation banner on user-facing docs.
+- **Track 2**: GitHub label scaffold (`migrated-to-agend-terminal`, `wontfix-deprecated`), retroactive labelling on this sprint's closed issues, `.github/ISSUE_TEMPLATE/config.yml` redirect link.
+
+Sprint 1 dispatch follows once this wrap doc is merged.