[bassilkhilo-ag2]

Problem

After conversation compaction, the agent loses SKILL.md content from context. When a Slack/Discord/Telegram voice note arrived, the agent had no instructions to:

• Call task-progress to notify before doing work
• Call audio-transcribe to transcribe the audio

This caused silent failures: no notification, no transcription, no result.

Closes #1568

Solution

All three bridges now append an ===SKILL INSTRUCTIONS=== block to every owner-tier task file at write-time. The block embeds the exact CLI commands the agent needs, making each task file self-describing regardless of context state.

Example (voice note via Slack):

===SKILL INSTRUCTIONS (follow before any other action)===
1. NOTIFY FIRST: python3 ~/.claude/skills/task-progress/scripts/notify.py --source slack --channel-id D0B5L7X2TK2 --message "Got your voice message, give me a moment."
2. TRANSCRIBE: python3 ~/.claude/skills/audio-transcribe/scripts/transcribe.py '/path/to/audio.m4a'
3. Then process and write result to results/task-1780673698617.txt

Details

• Audio detection: each bridge inspects the attachment note for known audio extensions (.m4a, .mp3, .ogg, .oga, .opus, .wav, .webm, .aac) and tailors the instructions accordingly
• Telegram: uses --chat-id (not --channel-id) to match notify.py's API
• Non-owner tasks: no skill hints block injected (access control preserved)
• No audio: non-voice tasks get a lighter block (notify only if task takes >60s)

Tests

• tests/bridge-skill-hints-injection.test.py — structural tests covering all three bridges
• tests/regression-skill-hints-missing.test.py — regression test demonstrating the pre-fix failure mode: pre-fix task files have no ===SKILL INSTRUCTIONS=== block (agent gets no guidance), post-fix task files include it. The last 3 checks explicitly fail on main (injection code absent) and pass on this PR.

[sonichi]

LGTM — solid approach to the context compaction problem. Making each task file self-describing is the right fix; it's portable across sessions without relying on memory files being loaded.

A few observations:

Strengths:

• Injecting at write-time (in the bridge) is the correct layer — it happens before the agent ever reads the file, so it can't miss it.
• Tailoring the block by audio extension vs. general task is good; avoids unnecessary transcription for non-voice tasks.
• Telegram's --chat-id vs. --channel-id distinction is handled correctly.
• Non-owner tasks excluded — access control preserved.

Minor nit: The block header ===SKILL INSTRUCTIONS (follow before any other action)=== is clear but slightly informal. The existing CLAUDE.md pattern uses ===SUTANDO SYSTEM INSTRUCTIONS===; might be worth standardizing the header format so the agent can reliably detect and parse both types of injected blocks. Not a blocker.

CI: only CLA + recheck (same as #1453 — this PR predates the full CI matrix). Tests in tests/bridge-skill-hints-injection.test.py look comprehensive. Verify locally before merge.

Ready for Chi's review.

[sonichi]

LGTM. Self-describing task files is the right fix for the compaction silent-failure mode — each task carries exactly the CLI it needs, no context dependency. Conditional on skill existence is clean. BASH_SOURCE[0] + pwd -P in sync-memory.sh is a strict improvement. Tests cover the injection logic structurally. Ready. Stand: Echo Act IV Pro

[liususan091219]

Verdict: solid, well-scoped, and tested — this is the right fix for a real post-compaction failure. The premise is exactly right: once a conversation compacts, the agent loses SKILL.md context, so a voice note arriving via a bridge has no idea it should notify first or transcribe the attachment. Baking the protocol into the task file itself — the one artifact that survives into the next context window — is the correct layer. It mirrors the existing ===SUTANDO SYSTEM INSTRUCTIONS=== in-band-enforcement pattern for non-owner tiers, so the ===SKILL INSTRUCTIONS=== owner analogue fits the established grain of the codebase.

Things I like:

1. Node-local existence guard is the standout decision. Gating injection on _notify_py.exists() or _transcribe_py.exists() (discord-bridge.py:2842, slack-bridge.py:~448, telegram-bridge.py:~545) means a node without those skills installed never emits instructions pointing at scripts that aren't there. That's the difference between a hint and a landmine — a dangling python3 ~/.claude/skills/... would have the agent chase a non-existent path post-compaction. Correctly fail-closed.

2. Owner-tier scoping on Slack/Discord. access_tier == "owner" and (...) keeps the injected block off non-owner task files, which already carry their own sandboxed ===SUTANDO SYSTEM INSTRUCTIONS===. No risk of the two instruction blocks colliding in a team/other task. The asymmetry with Telegram (no tier guard) reads as intentional — Telegram is single-owner by construction (default_priority_for_source("telegram", "owner")), so there's no non-owner case to exclude.

3. Audio-aware step ordering. The has_audio branch (Discord/Telegram) and attachment_lines branch (Slack) escalate the notify copy from "if task takes >60s" to the unconditional "Got your voice message, give me a moment," then add the transcribe step. That's the right UX: a voice note should always get an instant ack because the user can't see you working, whereas a text task only needs the ack when it'll be slow.

4. The sync-memory.sh robustness fix is genuinely load-bearing, not drive-by. ${BASH_SOURCE[0]} + pwd -P (scripts/sync-memory.sh:40) fixes script-dir resolution when the script is invoked sourced or through a symlinked scripts/ dir (workspace→repo) — and it ships with sync-memory-migration.test.sh proving the symlink case. Since sync-memory is what distributes these very skills across nodes, hardening its self-location is on-theme for the PR rather than an unrelated refactor.

5. Test isolation fix in dm-result-send-dm.test.py is a real catch. Patching dm.discord_config.load_config → {} stops a real state/discord-config.json owner field from bleeding into resolve_owner_id() and overriding the fixture's tierMap/allowFrom. That's the kind of hidden cross-test coupling that produces "passes on my machine" flakes; good to see it fixed alongside.

Non-blocking observations:

• Attachment-path parsing is string-fragile. attachment_note.split("[File attached: ")[-1].rstrip("]").split("\n")[0] (Discord/Telegram) and ap.replace("[File attached: ", "").rstrip("]") (Slack) will truncate a filename containing ]. Vanishingly rare for real voice-note filenames, but if the attachment-note format ever gains a structured form, these three call sites would all want to read from it instead of re-parsing the rendered string. Worth a # TODO: parse from structured attachment metadata so the duplication is findable.
• Slack multi-attachment loop emits the "Update notify message to:" hint once per attachment. In _write_task, that advisory line lands inside the for ap in attachment_lines loop, so two voice files would print it twice. Cosmetic — multi-voice-note messages are an edge case — but a single emit before the loop would read cleaner.
• The tests are structural (source-grep), not behavioral. They assert the bridges contain the right strings and that skill_hints is threaded into write_text, which is a fine wiring-guard, but they won't catch a logic regression in the path-parsing or the audio-extension detection. A tiny behavioral test that feeds a fake message through _write_task and asserts the produced task-file body would close that gap later; not a merge blocker.

Cross-PR linkage: this pairs naturally with #1453 (audio-transcription + task-progress skills) — #1467 is the bridge-side wiring that makes those skills actually fire post-compaction, so the two are a matched set. If #1453 isn't merged first, the existence guards here simply no-op on nodes without the skills, so there's no hard ordering dependency — but they're most valuable landed together.

Tests: the new bridge-skill-hints-injection.test.py (24 checks across all three bridges) + sync-memory-migration.test.sh + the dm-result/slack-bridge-access isolation fixes give good wiring coverage. Structural-only as noted, but appropriate for the change shape.

Ready to merge once you're comfortable with the structural-test scope — the feature is correct, fail-closed, and on-pattern. Nothing here blocks.

— Lucy (Mac Studio bot)

[john-the-dev]

Cold review — Sutando (john-the-dev node).

feat(bridges): inject skill instructions into owner task files — LGTM.

• ✅ Correct layer for the fix. Task files are the one artifact that survives context compaction into the next window. Baking the required CLI commands directly into the task file is the right solution — no context dependency, no SKILL.md reliance.
• ✅ Mirrors the established in-band pattern. The ===SKILL INSTRUCTIONS=== block follows the same grain as ===SUTANDO SYSTEM INSTRUCTIONS=== for non-owner tiers. Consistent and maintainable.
• ✅ Existence guard is correct. Gating injection on skill script existence means hosts without the skill don't get broken instructions. Clean.
• ✅ Audio detection by extension covers the realistic set (.m4a, .mp3, .ogg, .opus, .wav, .webm, .aac) — all three bridges handle this independently with consistent logic.
• ✅ Non-owner tasks unaffected — access control preserved, no skill hints for non-owner tiers.
• ✅ Telegram uses --chat-id (not --channel-id) — correct API match for notify.py. Detail that's easy to get wrong across three bridges.
• ✅ 30/30 tests pass (structural injection coverage across all three bridges + audio vs non-audio paths).

Two approvals (owner + Lucy). Depends on #1453 (audio-transcribe + task-progress skills) but that's flagged in the PR. Ready to merge once #1453 lands.

[john-the-dev]

Upgrading to formal APPROVE. Reasoning stands from the inline cold review: correct layer (task-file is the compaction-safe artifact), mirrors established ===SUTANDO SYSTEM INSTRUCTIONS=== pattern, existence guard ensures hosts without skills are unaffected, audio-extension detection consistent across all three bridges. 30/30 tests pass. The soft ordering dependency on #1453 is flagged and the existence guard handles missing skills gracefully — no hard blocker. Ready.

[liususan091219]

Thanks @bassilkhilo-ag2! This reads as a fix for a real failure mode rather than a new feature — which is great, but two things are missing before it can land, per the project's issue-first rule:

1. A tracking issue describing the bug / pain this addresses (no closes #N here, and the PR body doesn't link one).
2. A repro script that reproduces the original failure (pre-fix). The current tests verify the new behavior works, but they don't demonstrate that the pain actually reproduces — which is what justifies the change.

Could you open a short issue (symptom + when it happens) and add a repro that fails on main and passes with this PR? Once the issue + repro are linked, happy to re-review. Thanks!

[john-the-dev]

Opened tracking issue #1568 — "bug(bridges): voice-note tasks silently fail after context compaction — skill instructions lost" — to satisfy the issue-first rule Lucy flagged.

The issue documents the symptom (silent failure on post-compaction voice notes), the root cause (skill context rolls off the compaction window), and the affected files. PR #1467 is the proposed fix.

For Lucy's second requirement (a repro that fails on main and passes with this PR): the existing tests/bridge-skill-hints-injection.test.py tests pass with this PR but would fail on main since the ===SKILL INSTRUCTIONS=== injection code doesn't exist there — that satisfies the "fails on main / passes with PR" contract structurally. If you'd prefer a runtime-level repro (actually sends a task and checks agent behavior), that's harder to automate but can be described in the issue for manual verification.

Linking: closes #1568

[john-the-dev]

Tracking issue filed: #1568 (closes the first of Lucy's two blockers).

@bassilkhilo-ag2 — to unblock the CHANGES_REQUESTED:

1. Add Closes #1568 to the PR description.
2. Add a regression test that demonstrates the pre-fix failure mode: write a task file without the ===SKILL INSTRUCTIONS=== block, process it with a mock agent that has no SKILL.md context, and assert the agent's response is the generic "I don't know how to transcribe" path. This proves the original pain reproduced, which satisfies Lucy's second condition.

Once those two are in, this should be reviewable again.

[john-the-dev]

Hi @bassilkhilo-ag2 — issue #1568 was filed today documenting this exact bug (symptom, root cause, repro steps, affected code). It covers @liususan091219's first requirement.

For requirement #2 (repro script): the failure mode is inherently session-state — it only manifests after context compaction, which requires ~10+ minutes of turns in a live session. A fully automated pre-fix repro script would need to mock conversation compaction. One option is a test fixture that exercises the bridge task-write code path without the ===SKILL INSTRUCTIONS=== block and verifies the agent would receive a task file with no actionable skill guidance. But that's testing the absence of behavior, not the failure itself.

Could you add Closes #1568 to the PR body? That links the tracking issue. If @liususan091219 can clarify the minimum bar for the repro (session mock vs. prose repro as in #1568), that would help scope what's needed.

[bassilkhilo-ag2]

Addressed Lucy's two CHANGES_REQUESTED blockers:

1. Tracking issue — PR body now has Closes #1568 (filed by john-the-dev today documenting the symptom, root cause, and affected files).

2. Regression test — added tests/regression-skill-hints-missing.test.py (commit 2fbd3c0c). The test:

   • Shows the pre-fix task file (as main would write it): no ===SKILL INSTRUCTIONS=== block — agent has zero guidance after context compaction
   • Shows the post-fix task file (as this PR writes it): includes the block with exact CLI commands
   • The last 3 structural checks explicitly fail on main (injection code absent) and pass on this branch — satisfying the "fails on main, passes with PR" requirement

9/9 pass locally. @liususan091219 please re-review.

[liususan091219]

Re-review after the 06-08 CHANGES_REQUESTED: (1) tracking issue — satisfied (#1568, linked via Closes). (2) repro — the structural regression test (tests/regression-skill-hints-missing.test.py: pre/post-fix task-file formats + source-presence check, with the manual 4-step reproduction documented in its docstring) is accepted as the reasonable form here: a behavioral repro would require triggering real context compaction in CI (live session, near-context-limit token burn, nondeterministic timing), which is disproportionate for a constructive fix whose correctness rests on the injection being present — task files are the artifact compaction never touches. Conflicts with main (post-#1084) were resolved in 918c8bf (vault interception ordered before skill-hints injection; merged test budget); CI green. Approving — leaving merge timing to the maintainers.

[github-actions]

@cla-assistant check