For measurable task classification and upgrade rules, see management intensity. The frozen first-batch assessment envelope and reuse mapping are in engineering assessment.
Use AWR's source ledger as a small work contract. Keep the outcome, current state, completion criteria and next action there. Link design documents and detailed reports instead of duplicating their contents. Chinese text and paths are supported; short writing is a convenience, not a parser requirement.
awr work create --input draft.json previews one draft containing explicitly known facts. Apply that same input with --accept --expected-preview <fingerprint> --expected-revision <revision>. The stable request key identifies the whole input. Query work create-status --key … after an uncertain response; use the existing explicit recovery command for a pending outcome. Do not generate another key to retry.
{
"version": 1,
"request_key": "review-invoice-2026-09",
"title": "核对本月发票",
"fields": {
"summary": "核对金额与订单记录",
"acceptance": ["差异清单可供复核"],
"next_action": "读取已授权的发票和订单"
}
}
Creation always produces a draft. Missing goals or required facts remain visible; creation does not grant permission to execute or complete. Supported fields are summary, goal, acceptance, next_action, kind, paths, tags, priority, depends_on, owner and milestone. Markdown sources must support the requested fields in their existing table/checklist shape; unsupported shapes fail before writing.
awr work prepare WORK --session SESSION --source-sha FULL_SHA returns readiness, the required context, claims, persistent waits and a next action from one snapshot. Use --goal for an explicit goal selection. The MCP equivalent is awr_work_prepare (also accepts a bound conversation). The caller must actually consume the context before acknowledging it in a checkpoint. Readiness is not a completion claim. CLI refreshes the projection; MCP verifies sources without writing.
awr work prepare-completion WORK --report report.json --evidence-key KEY --source-sha FULL_SHA --level locally_verified reads actual report bytes and returns the evidence arguments and the acceptance mapping for the existing completion operation. awr_completion_prepare provides the same preflight through MCP. No command runs, no evidence is registered and no task is completed by preflight. The level is the caller's declaration. The command, scope and time come from the report; the file digest is computed by AWR. The supplied source SHA must match the report and must include any changes relevant to verification in its accompanying source evidence. AWR does not assume a clean Git checkout.
The report uses this structure; replace every fixture value with actual evidence:
{
"version": 1,
"work_item": "WORK",
"source_sha": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
"command": "the command or verification procedure actually executed",
"scope": ["the scope actually verified"],
"verified_at": 1,
"checks": [{
"name": "independent check name",
"passed": true,
"details": "the observed outcome, not an intended result",
"criteria": ["an exact current source acceptance criterion"]
}]
}
Preflight checks all current criteria. Recording and completion remain separate writes with current revisions. Completion rechecks claims, dependencies, source freshness and report bytes; changing a report after preflight invalidates its digest. Query a shared MCP write by its stable request ID when its outcome is unknown.
When status is checked with a source SHA, an unchanged task completed through AWR is verified against the evidence selected by that successful completion on the same runtime branch. Every selected report, including extra required evidence, is read and checked again; another passing report cannot replace a missing or damaged selected report. Rejected attempts remain available in evidence history. If the task facts/revision, branch or requested source SHA no longer match the completion receipt, status uses the existing conservative source/evidence assessment. Source-declared completion alone never supplies a verified receipt.
For one preparation, the old work-get plus context-compile pair becomes one call with the same required context. This is a call-count reduction only, not a claim about net model tokens, billing or total task duration. Measure those across the whole workflow, including maintenance and recovery.
