cogpros Closing Time Buzz Facts
Close the Buzz working day across channels, threads, and agents. Ask appraisal questions in the terminal. All relay access is read-only: no messages, reactions, canvas updates, or presence changes. Reports, measurement, and seals stay local.
Setup
Requires Python 3.10+, bash, and a Buzz CLI configured with a read-only identity.
Configure credentials through your normal secret manager; never paste them into
this skill or print them. Set BUZZ_RELAY_URL to your community and
BUZZ_OPERATOR_NAME to the operator's exact display name, or use
BUZZ_OPERATOR_PUBKEY for an unambiguous identity. The collector uses
existing CLI authentication; it does not select another identity on failure.
State defaults to ${CLOSING_TIME_STATE:-$HOME/.closing-time/state}/closing-time-buzz.
BUZZ_WORK_LOGS_DIR optionally supplies the work-log directory. The bundled
Ghost Hours writer respects GHOST_HOURS_LOG and its normal configuration.
Keep the repository intact so the shared writer and scoring references resolve.
Capture the working day
Read ../closing-time/references/scoring-constants.md, then privately compute
and retain the agent's blind FW-C. Do not reveal it during appraisal.
Run the bundled collector and save its JSON output locally:
bash <skill-dir>/scripts/buzz-stack-facts.sh --json > <capture.json>
Without --json, the same collector prints a readable fact sheet. For an
explicit calendar-day backfill, add --date YYYY-MM-DD. Normally the window runs
from the last seal to now, including work after midnight. With no seal within
48 hours, the collector uses the current calendar day and labels that fallback.
Keep the first successful capture fixed for this close; do not shift the window
while answering appraisal questions.
An unreadable channel, unresolved writer, or truncated message page blocks a complete capture. Explain the missing evidence and leave the day unsealed. Do not treat quiet agents as dead or treat unavailable data as zero work.
Read and synthesize
Deduplicate shipped outcomes across threads and available work logs. Surface contradictory accounts, dangling threads, and stale or missing canvases. Thread silence is a prompt for judgment, not proof that the work failed. Root IDs and timestamps remain attached to the evidence.
Carry forward real loose ends in your existing task system, using stable keys and enough context to resume: state, attempts, files, first command, done condition, and source session. Do not turn closing into unrelated implementation.
Appraise once
Human and agent time come from the captured message timestamps, with the collector's 30-minute gap cap. These are modeled timing estimates, not direct observation of attention. GH is a separate counterfactual estimate for the shipped work; use the shared taxonomy and type-specific anchors.
Ask one question at a time:
- FW-C 1–10 for the Buzz working day.
- If FW-C >= 5, request an optional verbatim note about why it mattered.
- Give a GH range and counterfactual explanation; let the operator choose.
If the operator explicitly delegates the appraisal too, estimate the proxy
FW-C and mark fwc_source=agent-blind with operator-override-fill. Keep the
original blind score separate. Do not fabricate a quote; omit it when absent.
Use operator provenance only for an operator-supplied score.
Record and seal
Render a local HTML report with shipped work first, followed by the thread table, roster, timing and appraisal, loose ends, and canvas findings. Do not post the report to Buzz. Append a short close narrative to your daily notes.
Save appraisal JSON, with values from the current close:
{"type":"speed","gh_mins":120,"fwc":5,"fwc_source":"operator",
"fwc_eom":4,"desc":"Completed the day's agreed work","tags":[]}
Optional fields: subtype, note, and backlog_months. A note must be the
operator's actual words. Then write the measurement and seal:
python3 <skill-dir>/scripts/record-close.py --capture <capture.json> \
--appraisal <appraisal.json> --report <report.html> --seat <agent-label>
The helper preserves the buzz-stack-YYYY-MM-DD join key, uses source
buzz-stack, and verifies the canonical measurement before sealing. Matching
retries reuse the record. Changed re-closes require an explicit amendment;
never silently append a duplicate day or rewrite the historical measurement.
Return the report link, captured window, HH/GH, FW-C with its provenance, remaining work, and seal result. Keep the blind score private. A successful report alone is not a successful close if the measurement or seal failed.
Limits
This public skill closes the whole working day; per-thread close is not exposed by its record helper. Overlapping threads can double-count attention, and gaps are capped. A page at the configured retrieval limit is treated as incomplete. The collector does not diagnose runtime health or fill missing relay evidence.