alfred-hours-reconstruction

name	alfred-hours-reconstruction
description	Reconstruct Sir's work on a matter from bounded cross-source evidence — calendar, messages, mail, commits, transcripts, sessions, artifacts — and produce an auditable per-period proposal for his approval. Never accepts hours automatically. Use when a tracked period closes, when Sir asks "how much did I work on X?", or for a one-off retrospective backfill. Activates for matters carrying `hours_tracking` in their frontmatter.
version	1.0
metadata	{"openclaw":{"emoji":"⏱️"}}

Alfred — Hours reconstruction

Reconstruct what Sir actually worked on, from evidence, and propose it.

Two rules sit above everything else here:

Nothing enters the ledger without Sir's explicit approval. The accepted ledger is the cursor — not the calendar, not the schedule.

This is an evidence-based estimate. It is not a timer export, not an invoice, and not automatically accepted time.

Switching it on

hours_tracking: true

Off by default. Everything is derived: the period is weekly, closing at the end of Sir's Friday in his timezone; the approval request goes wherever Alfred normally reaches him; the ledger is note/<matter-slug>-timesheet.md and the proposals are note/<matter-slug>-timesheet-proposal-<period-end>.md.

Object form only to override:

hours_tracking:
  enabled: true
  period: weekly
  approval_channel: <channel>
  rate: 125            # optional, for later invoicing; not needed to track

Same boolean-or-object shape as commitment_register. A capability is a shipped skill, switched on by a boolean on the matter, with an object form as the escape hatch.

The window comes from the ledger

The window runs from the last date covered by the accepted ledger (exclusive) to this period's cutoff (inclusive).

Not from the calendar, and not from the last time the schedule fired. That distinction is the whole point: if a run is missed, the next window widens to cover the gap. Keyed to the schedule instead, a missed Friday silently loses a week of Sir's work, and nothing ever reports it.

Before reconstructing anything, read the existing ledger and record its latest covered date, closing balance, credits and rate. Reconstruct only the uncovered period — never lay a partial reconstruction over a period a fuller ledger already represents.

Weekend work therefore carries into the following period rather than vanishing.

Gather bounded evidence

Read only the window. Keep a source handle for everything.

1. Calendar — every owned calendar, not just the primary. Count meetings at actual start/end. Corroborate with transcripts or meeting notes where available.
2. Messages — Sir-authored messages, threads, attachments and file deliveries in matter-related channels and DMs. Include replies whose own timestamps are in-window even when the thread parent is older.
3. Mail — sent and received threads in the window. Metadata first, then fetch only relevant threads. Reading an email is not automatically work.
4. Transcripts and meeting notes — to classify discussion, preparation, follow-up and deliverables. Transcript length is not duration.
5. Repositories — commits, PRs, reviews, issues, releases attributable to the matter. Commits prove activity and output, not duration.
6. Sessions — Sir's own conversations with Alfred about this work. Exclude register refreshes, cron maintenance and repeated summaries unless he was actively directing substantive work.
7. Vault and artifacts — record transitions, generated files, delivery evidence. Timestamps anchor work; they do not determine hours.

If a configured source is unavailable, say so and lower confidence. Never compensate by inventing time.

Then convert evidence to hours by the method in references/evidence-to-hours.md. The union-of-intervals rule is not optional: the same hour seen in four sources is one hour.

Write the proposal — never the ledger

Search first, so one period has exactly one proposal. Write note/<matter-slug>-timesheet-proposal-<period-end>.md with:

status: draft
accepted: false
period_start: <date>
period_end: <date>
generated_at: <ISO timestamp>
total_hours: <number>
scope: <matter-slug>
source_coverage: <which sources were read, which were unavailable>
confidence: <overall>

status must be draft, not proposed. A proposal is a note, and the vault validator accepts only active|draft|final|review there. An invalid value written here is inert at creation — the raw-content path skips validation — but the validator re-checks the whole record on the first PATCH, so it then blocks the acceptance and every later write to that record. Acceptance sets final. accepted: true/false is the machine marker either way.

Body:

1. Daily table — date, proposed hours, short work summary, linked commitment IDs where a register exists, per-day confidence.
2. Evidence table — time, source, exact handle, activity, interval used.
3. Exclusions and uncertainties.
4. Total and its effect on capacity, clearly labelled provisional. Never present unapproved excess as billable — see ../alfred-commitment-register/references/contract-authority.md.

Read the note back before delivering anything.

The ledger is not touched. Not on generation, not on a test run, not because the reconstruction was thorough.

Ask for approval — on the Desk

A proposal delivered only as a message is a message: easy to miss, easy to defer, invisible once scrolled past. Write a Desk card, so the decision sits where Sir's other decisions sit.

Create a needs_attention record carrying:

approval_kind: hours_proposal
proposal_ref: note/<matter-slug>-timesheet-proposal-<period-end>.md
ledger_ref:   note/<matter-slug>-timesheet.md
action_what:  "Approve <n.nn> h for <matter name> — <period-start> to <period-end>"
suggested_actor: human

Those three fields are a contract, not decoration. route_decision reads approval_kind to recognise the card, and proposal_ref / ledger_ref to know what to book. A card missing any of them books nothing — the click will close the card and silently do nothing else.

Then deliver the itemised request as well: the period and total; one line per non-zero day; the confidence and coverage caveat; the proposal note's path; and an explicit prompt to approve or supply a corrected figure.

Tell Sir he can correct the number in the note when approving — typing 6, or 6.5 — Tuesday ran short, books that figure instead of the estimate. A note that is only prose (looks fine) approves the estimate unchanged; the number is never guessed from wording.

Only Sir's explicit approval of the identified period is acceptance. A generated proposal is not approval. Silence is not approval. A successful scheduled run is not approval. A reaction or a "thanks" is not approval.

On approval

Approving the Desk card does the booking automatically — route_decision appends the period to the ledger, then marks the proposal accepted, then reads both back. Do not perform these writes yourself when the card path was used; you would double-append, and the ledger is the cursor for the next window, so a duplicate corrupts the following period too.

When approval arrives some other way — Sir replies in chat, or no card was written — do it by hand, in this order:

1. Read the exact proposal note. Confirm its period overlaps no other accepted or proposed record.
2. Patch it to status: accepted, accepted: true, accepted_at, accepted_by. Preserve the original evidence.
3. Append the accepted period to the ledger — which advances the cursor for the next window.
4. Read both back and verify the period appears exactly once and the totals agree.

On correction

When Sir supplies different hours instead of approving, record both his figure and the original estimate, with the delta.

That delta is the only feedback signal in this whole feature. Without it the estimator is permanently as good as it was on day one — and it is the thing most easily dropped as a nicety, because storing only the correct number obviously "works".

Then patch the proposal, recompute totals, read it back, and repost the corrected proposal. Do not silently accept a corrected figure.

Retrospective backfill

Reconstructing history is a different job from closing a period: a broad sweep rather than a bounded one, and it should stay manual and on demand. It is not part of the scheduled run.

When Sir asks for one, establish the contract first — inclusive dates, timezone, opening balance, credits and their effective dates, whether the estimate should be conservative, central or generous, and the matter's aliases. Preserve his exact ordering when he gives a rule such as "start at zero on the 18th, then add 40 hours on the 19th":

closing = opening + credits effective that day − estimated work that day

For long ranges, produce a Markdown or CSV artifact with date, estimated hours, concise work, credit and closing balance, plus a methodology section. Read it back and verify the totals, end date and final balance.

If Sir first asks for a generous estimate and then a conservative commercial presentation, re-estimate the daily rows — do not scale the headline. Reduce uncertain prep, editorial and follow-up time before touching exact meetings or directly evidenced work. A requested ceiling is a scenario constraint, not permission to fabricate contemporaneous time; if it cannot be met without contradicting exact evidence, surface the conflict rather than forcing the ledger.

Failure behaviour

• One optional source fails → continue, and label source coverage degraded.
• Delivery fails after the proposal is saved → do not create a duplicate note. Retry delivery only.
• Vault access fails → do not post a proposal as though it were durable.
• Never send client-facing messages, invoices or timesheets from this workflow.

Pitfalls

• Do not sweep unbounded history when the period is known.
• Do not trust one keyword; a generic product name matches unrelated threads.
• Do not count automated jobs, digests or dashboard refreshes as Sir's labour.
• Do not silently convert an estimate into an authoritative timesheet.
• Do not infer billability or client debt from a negative balance — resolve commercial treatment from the executed agreement first.
• Do not hide uncertainty. Label the method and the confidence.
• Do not omit zero days.
• Do not accept hours at any confidence level, ever.