User pushback: "что теперь делает сон? все связано?" — Sleep Phase B
was reading only `traces/`, ignoring the four tracking journals shipped
in the previous commit. Cloud agent had a partial view of what happened.
This commit closes the loop. Sleep now sees everything that's tracked.
PUSH SIDE — `kei-sleep-sync.sh` (called on every Stop event)
Now mirrors the full observability surface into the memory-repo:
~/.claude/memory/time-metrics/sessions.jsonl → time-metrics/
~/.claude/memory/time-metrics/tasks.jsonl → time-metrics/
~/.claude/memory/time-metrics/numeric-claims.jsonl → time-metrics/
~/.claude/memory/time-metrics/agent-toolstats.jsonl→ time-metrics/
~/.claude/agents/ledger.sqlite agents table → ledger/agents.jsonl
~/.claude/agents/ledger.sqlite skill_invocations → ledger/skill_invocations.jsonl
Format: JSONL (one row per object). The two ledger tables are dumped
via `sqlite3 + json_object()` so cloud agents can stream-parse into
pandas / duckdb without binary-file handling.
First sync moved 6 files / 638 rows from local to remote — verified
by `git show --stat` of the resulting `memory: session traces` commit.
CONSUME SIDE — `phase-b-rem.sh` REM-consolidation report
Each nightly `reports/sleep-YYYY-MM-DD.md` now ends with a "Tracking
observability (last 7 days)" section containing four jq-aggregated
digests:
1. Agent outcomes — per-model: n, functional/partial/scaffolding/fail
counts + total_cost_usd. Lets the agent see whether the model-tier
refactor (cb1fdde) actually paid off and whether Sonnet success
rate justifies routing more task classes to it.
2. Skill success rates — per-skill: n, successes, rate_pct. Drives
Phase D nightly decisions (archive unused / re-extract failing /
mark validated). Empty until Skill tool is invoked in the next
session.
3. Numeric-claims tier breakdown — REAL / FROM-JOURNAL / ESTIMATE-HTC
counts. High ESTIMATE-HTC ratio = orchestrator under-calibrated.
Cloud agent's job: spot frequent ESTIMATE-HTC categories and
propose conversion to FROM-JOURNAL via measured runs.
4. Agent tool-call patterns — mean tool_use_count, mean duration_ms,
per-tool total calls. Lets the agent see "this code-implementer
spawn made 30 Read but 1 Edit — was tier-allocation correct?".
All four sections gracefully skip if the source JSONL is missing or
empty. jq is the only new dependency (already present per existing
phase-b checks).
What is NOT yet automated:
- The cloud agent's prompt template doesn't yet INSTRUCT it to act
on these digests. Currently the digest is data; whether the agent
proposes rule + hook codification based on it depends on the
free-text instructions in the schedule. Follow-up: codify a Phase B
instruction block that maps each digest to a recommendation pattern.
- Idempotency on `cp` for time-metrics: I use plain `cp` (not `cp -n`)
so the latest local state always overwrites remote. The journals are
append-only on the local side, so this is safe — but if two machines
ever share one memory-repo it would corrupt. Out of scope for
single-machine setup.
=== STATUS-TRUTH MARKER ===
shipped: functional
stubs: 0
cargo-check: NOT-RUN (pure shell)
behaviour-verified: yes
follow-up-required:
- Phase B prompt template — instruct cloud agent to act on the four
digests (codify recurring patterns, calibrate ESTIMATE-HTC).
- skill_invocations.jsonl will populate from next session onward.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the loop on "without full tracking the system can't make decisions"
(user pushback on partial coverage). Three gaps that left the inference
layer blind are now wired:
GAP #1 — agent toolStats / token counts / cache hits captured
================================================================
`agent-outcome-backfill.sh` now appends one JSONL row per spawn to
`~/.claude/memory/time-metrics/agent-toolstats.jsonl` with:
agent_id, outcome, stubs, ts,
tool_use_count, duration_ms, tool_stats {Read:N, Bash:M, ...},
tokens_in, tokens_out, cache_read, cache_write
Sidecar journal (no schema migration). Production payload's
.tool_response.totalToolUseCount / totalDurationMs / toolStats / usage
fields land directly. Smoke-tested with synthetic spawn — row written.
GAP #2 — skill_invocations table actually receives writes
================================================================
The `skill_invocations` table (schema v8) had 0 rows because no caller
existed for `skill_metrics::record_invocation`. Added two pieces:
(a) `kei-ledger record-skill <name> --success {0|1}` CLI subcommand
Mirrors record-cost; same dispatch shape. Optional `--agent-id`,
`--trajectory-id`, `--duration-ms`, `--db`. Validates non-empty
name + duration ≥ 0. Outputs `{"ok":true,"skill":"...","ts":N}`.
(b) `hooks/skill-record.sh` — PostToolUse:Skill hook. 50 LOC POSIX.
Detects Skill tool calls, derives success heuristic from
tool_response (exit_code / status / content non-empty), shells
out to `kei-ledger record-skill`. Bypass via SKILL_RECORD_BYPASS=1.
83 kei-ledger tests pass (16 unit + 67 integration). Smoke-tested
end-to-end: `kei-ledger record-skill test-skill --success 1` inserts
a row with correct fields.
Phase D nightly skill-metrics decisions (archive if unused N days,
re-extract if success<60% over M days, validated if >20 calls + >90%
success) now have data to consume.
GAP #3 — numeric-claims.jsonl receives every evidence-tagged claim
================================================================
RULE 0.18 mandated three markers `[REAL:]` / `[FROM-JOURNAL:]` /
`[ESTIMATE-HTC:]` on every numeric/duration/cost claim, but no hook
appended valid claims to the journal — the calibration data RULE 0.18
promised never accumulated.
`hooks/numeric-claims-record.sh` — Stop hook, 140 LOC POSIX. Reads
transcript_path from stdin, locates the last assistant message via
recursive flatten (same pattern as agent-outcome-backfill.sh after
the production-payload-shape fix), regex-extracts every `<phrase>
[<TIER>: <pointer>]` triple, appends one JSONL row per claim.
Idempotent within 1-second window to avoid double-recording on
repeat Stop fires. Bypass via NUMERIC_CLAIMS_RECORD_BYPASS=1.
Smoke test: synthetic transcript with 3 markers (REAL + ESTIMATE-HTC
+ FROM-JOURNAL) produced exactly 3 well-formed JSONL rows.
Settings.json
================================================================
- PostToolUse:Skill matcher created (or augmented if already
present) with skill-record.sh.
- Stop:* matcher gains numeric-claims-record.sh after the existing
chain (stop-verify, task-timer, session-end-dump, extract-task-
durations, chat-numeric-postflag, affect-threshold-check,
enrich-from-jsonl).
What this does NOT do (deferred):
- Backfill `skill_invocations` from past traces (history started
today; Phase D cohort builds forward from now).
- Migrate the agent toolStats sidecar JSONL into a proper ledger
column. Append-only file is fine for the current scale.
- Refactor main.rs (now 233 LOC, was 212; pre-existing CP debt
flagged by skill-record agent — separate cleanup PR).
=== STATUS-TRUTH MARKER ===
shipped: functional
stubs: 0
cargo-check: PASS
behaviour-verified: yes
follow-up-required:
- kei-ledger main.rs Constructor Pattern split (212→233 LOC)
- Verify in next session: skill_invocations gets rows from real
Skill tool use; numeric-claims.jsonl gets rows from real assistant
messages with markers
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hook never fired in production despite passing unit tests. Diagnosed
via debug-log + payload dump: real Claude Code PostToolUse:Agent sends
`tool_response` as an OBJECT (not string, not array), with the agent's
reply at `tool_response.content[0].text` — keys: agentId / agentType /
content / prompt / status / toolStats / totalDurationMs / totalTokens
/ totalToolUseCount / usage.
Original jq filter handled string + object (`$r.content // $r.text`)
but `$r.content` returns the array verbatim; `jq -r` then dumps the
JSON literal which has `\n` as escape sequences, defeating the
`grep -m1 '^shipped:'` line-anchor.
Fix: recursive `flatten` jq function:
string → as-is
array of any → recurse, join "\n"
object with .text → return .text
object with .content → recurse into content
anything else → ""
Verified end-to-end: latest 4 code-implementer spawns now write
outcome=functional to ledger correctly. Beta posterior in
kei-model-router begins receiving signal.
Production cleanup:
- Removed verbose debug-log + payload-dump diagnostic. Toggle via
`AGENT_OUTCOME_DEBUG=1` env if hook stops firing in some future
Claude Code version.
- Hook source committed to `hooks/agent-outcome-backfill.sh` so
`install.sh` deploys it on fresh installs (was only in user-home
previously — gap from `feat/substrate-path-atoms` agent run).
=== STATUS-TRUTH MARKER ===
shipped: functional
stubs: 0
cargo-check: NOT-RUN
behaviour-verified: yes
follow-up-required:
- none
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Parfii-bot