Cavan-Ou/dsh-observation-journal

Zero-touch runtime telemetry for DeepSeek Harness: every session writes its own report card.

简体中文版见 README.zh.md

A pure observer plugin (zero tools registered, zero LLM calls, zero agent involvement). When a session ends, it writes the run's facts — task, model tier, tools, failures, duration, status — into a human-readable journal with an auto-updated stats section.

Why it exists

Failures have a recorder (dsh-fail-logger). Successes and run facts didn't. This is the sibling: what happened — not how to solve it, not what to remember. No tools, no injection, no retrieval. The harness writes passively; humans and projects read.

What it is not (boundaries, stated plainly):

Plugin	Records	Injects back into agent?	Consumption
this	run facts (telemetry)	never	human/project file
dsh-task-planner	solutions ("how to solve")	yes (recall)	agent planning
dsh-mneme / dsh-memento / dsh-memory	agent memories	yes (retrieval)	agent context
dsh-fail-logger	failures	via skill	agent skill loading

60-second verification

dsh plugin --profile headless add <repo-or-pkg>   # or copy the repo as a local bundle
dsh --profile headless "run any small task"
cat ~/.dsh/observations.md                        # a journal row + stats section appeared

What the output looks like

The journal is the UI. A marker section that survives manual edits, plus an auto-stats block:

<!-- OBS-JOURNAL:BEGIN -->
| time | sid | task | model | dur | turns | tools | calls | fail | status |
|---|---|---|---|---|---|---|---|---|---|
| 2026-08-14T21:42 | abe96e0f | 阅读 specs/s11-1.md 任务书 | deepseek-v4-pro(max) | 1242 | 1 | read:80,bash:12,edit:9,todo_write:4 | 106 | 0 | completed |
| 2026-08-15T03:04 | 9c1f3a | run any small task | deepseek-v4-flash(max) | 25 | 1 | bash:4,grep:2,glob:1 | 7 | 0 | completed |
<!-- OBS-JOURNAL:END -->

<!-- OBS-JOURNAL:STATS -->
- sessions: 2
- failure rate: 0.0%（0/113）
- top tools: read:80,bash:12,edit:9,todo_write:4,grep:2
- avg duration (s) by model: deepseek-v4-flash(max): 25, deepseek-v4-pro(max): 1242
<!-- OBS-JOURNAL:STATS:END -->

raw sidecar (obsFile + '.jsonl', append-only): full fidelity — todo planning trace (≤5), complete tool counts, failed tools, full model id, full task description, normalized task_hash. This is the v2 material for LLM insight; it is TTL-decoupled from the card.

Config (all optional, patch config: field)

key	default	description
obsFile	$DSH_HOME/observations.md	journal path (point it at a project-level file)
maxRows	200	card keeps last N rows (raw sidecar unaffected)
marker	OBS-JOURNAL	section marker id, [A-Za-z0-9-]
redact	[]	extra redaction regexes (stacked on the built-in secret table)
flushMs	300	trailing debounce after turn/end

Env: OBS_FILE overrides obsFile; OBS_REPLAY=<session.jsonl> replays real events (test/CI mode).

Reliability

• 10-column card rows: one row per session — no lossy merging
• Task title escapes | and newlines; secrets redacted (same table as fail-logger)
• Cross-process write lock + stale lock reclaim; dispose fallback flushes sessions with no turn/end
• Tested against real session logs: 14/14 replay tests on 5 real .zstd fixtures (incl. a 2000+ event Pro long-synthesis session), field-by-field cross-checked against independent recomputation; 21-session full replay verified human sections byte-identical

Development

node --check lib/index.js
node --test tests/test.mjs   # needs python3 + zstandard

License

MIT