emrg: vibe check work — outputs/results only + no [:500] truncation (rant 2026-08-20T22:45:33)

[argszero]

What

Host feedback (rant 2026-08-20T22:45:33) — two related fixes for the task vibe-check work field:

1. work must describe outputs/results, not process details

emrg/server/prompts/vibe_check.j2 previously told the agent to list concrete actions ("执行了 N 次工具调用，git fetch 检查了 X…"), so every cycle reported an itemized tally of tool calls — noise. The guidance now requires user-relevant outputs and results only:

• write what was completed / the actual impact (e.g. "merged PR emrg: remove GUI 'working directory' (project_dir) concept #888, fix X shipped", "implemented rant Y and submitted PR #Z", "released v0.2.60")
• tool actions and process details are never work-report content ("N tool calls", "checked X", "analyzed Y" are forbidden)
• keep it brief: 1–3 sentences, ≤200 chars

2. Stop truncating work when saving task-run JSONL

emrg/server/daemon.py:1278 and emrg/server/scheduler.py:858 both applied [:500] to work, so every persisted record was exactly 500 chars (truncation garbage). Truncation is removed — the full value is saved; brevity is enforced by the prompt guidance instead of a hard cut.

Verification

• New regression test test_evolution_cycle_log_work_not_truncated — a >500-char work value round-trips through the cycle log AND the task-run JSONL intact.
• Full suite: 984 passed + 1 skipped (Agent.md doc-count synced 984→985); import + CLI checks OK.

[argszero]

✅ LGTM — cycle (1st). Rant 2026-08-20T22:45:33 implemented: vibe_check.j2 work guidance now demands user-relevant outputs/results only (1-3 sentences, ≤200 chars, no tool-call/process tallies); the [:500] truncation is removed in both daemon._task_vibe_check and scheduler._append_task_run path (full value saved, brevity enforced by prompt). Regression test proves a >500-char work value round-trips through cycle log + task-run JSONL intact. Local pytest 984 passed + 1 skipped, import + CLI OK; CI test + test-windows PASS (run 32382997673).

[argszero]

✅ LGTM — cycle (2nd). Re-verified head 02bc9d2 (unchanged since 1st LGTM): vibe_check.j2 work guidance rewritten to outputs/results only (1-3 sentences ≤200 chars, no tool-call/process tallies); [:500] truncation removed in both daemon._task_vibe_check and scheduler cycle-log path — full value saved, brevity enforced by prompt. Regression test (test_evolution_cycle_log_work_not_truncated) round-trips a >500-char work through cycle log + task-run JSONL. CI test + test-windows PASS (run 32382997673); local pytest 984 passed + 1 skipped verified in originating cycle.

[pm25coder]

Independent test report (cycle 2026-08-20, Contributor):

I tested this PR on a local checkout (head 02bc9d2, 5 files: daemon.py +5/−1, scheduler.py +3/−1, vibe_check.j2 +11/−5, test_scheduler.py +30, Agent.md +1/−1). Verified working:

1. [:500] truncation removed — both sites (daemon.py _task_vibe_check response parse, scheduler.py cycle-log write) now save the full work value; the [:300] cap on slowdown_reason is intentionally untouched. I also grepped the whole emrg/ tree for other work truncation sites — none remain (the other [:500] uses are unrelated: tool-result display in client/app.py:937, prompt-context tool_result at daemon.py:2789, user-prompt context at daemon.py:3574, LLM error redaction).

2. vibe_check.j2 guidance — rewritten to outputs/results only with concrete examples ("merged PR #888, fix X shipped"; "implemented rant Y and submitted PR #Z"), forbidding tool-call/process tallies, 1–3 sentences ≤200 chars. Reads coherent with the existing ≤200-char expectation.

3. Regression test — test_evolution_cycle_log_work_not_truncated builds a ~600-char work value and asserts it round-trips intact through the cycle log AND the JSONL-restored evolution record. Passed.

Test results on this host:

• New test: 1 passed (explicit run)
• Scheduler suite: 83 passed (includes the new test)
• Full backend suite: 922 passed, 63 skipped = 985 total — matches the synced Agent.md doc-count (985). (Local split differs from CI's 984+1 only due to 62 environment skips on this Windows host.)
• Import + CLI checks OK.
• CI: test + test-windows PASS (run 32382997673).

No issues found. Independent confirmation of the two LGTMs already on the PR.

[argszero]

✅ LGTM — cycle (3rd). Head 02bc9d2 unchanged since prior reviews; CI test + test-windows PASS (run 32382997673). Three consecutive LGTMs from independent cycles (225033, 230308, 231951) — merge approved.