Project Memory, Trust Report, and Outcome Mining

Yardlet v0.8 adds three loop-level surfaces that sit above the worker CLI: Project Memory, Trust Report, and Outcome Mining. They share one boundary: workers can use the context and propose improvements, but Yardlet keeps the canonical state and policy changes deterministic and auditable.

Project Memory

Project Memory is repo-local durable context. A workspace can keep facts and decisions as Markdown files under .agents/memory/, one fact per file. Each file may use optional frontmatter such as:

---
name: Payment webhook decision
description: Reservation finalization must be idempotent.
look_at:
  - src/webhooks.rs
  - docs/payments.md
---

Yardlet discovers .agents/memory/*.md during the same harness discovery pass that finds rules and skills. It skips the scaffold README and non-Markdown files. For every memory doc, Yardlet extracts only an index line:

• title, from name: / title: or the first # heading
• one-line summary, from description: / summary: or the first prose line
• anchor path, such as .agents/memory/payment-webhook.md
• optional look_at landmarks

That index is injected into the planner and every worker packet. The body is not inlined. The packet tells the worker to read the entry only when it bears on the task. This is the index-and-anchor model: always load a small map, then open the few bodies needed for the current work.

yardlet memory renders the same index for the human. If a memory file declares look_at: landmarks, the command checks git state for those paths. A memory is marked possibly stale when a landmark has an uncommitted change or a newer git commit time than the memory doc. The packet hot path stays git-free; staleness is a diagnostic view in the command.

yardlet memory init and yardlet memory refresh maintain the docs through a worker without hand-editing. init asks a worker to draft memory documents from the repo into an isolated run directory (memory-result.json); Yardlet's core is the single writer that turns those drafts into canonical .agents/memory/*.md. refresh re-drafts the existing docs the same way, and refresh --stale-only limits the worker to the docs currently flagged possibly stale. The worker only proposes content; the deterministic core owns every write.

Trust Report

The Trust Report is a deterministic read over your own history. It answers a local question: how much should this workspace trust the loop's Done outcomes? yardlet trust folds two sources, run telemetry (.agents/telemetry/runs.jsonl) and the state-transition audit log (.agents/transitions/<task>.yaml), into two layers, and only reports.

Attempt view (from run telemetry)

• first-pass Done, for tasks whose first recorded attempt reached Done
• Done after retry, for tasks that reached Done only after more than one attempt
• no Done in record, for tasks that have no Done telemetry row
• per-worker reliability, including done rate, partial count, failed count, no-result count, wall time, and user override count
• tasks that needed multiple attempts, sorted by most attempts first

Telemetry records carry intent_id. When the active intent has matching telemetry, the attempt view scopes to that intent so reused task ids do not fold together across unrelated intents. If there is no intent-scoped telemetry yet, it falls back to the cumulative view and labels that caveat.

Autonomy view (from the transition audit log)

The transition log is the only record where a Done that was later reopened, or a human intervention, is visible. Yardlet folds it, keyed per (intent, task) instance, into:

• Can I trust a Done? Each Done is graded as evidence-backed (a clean Done, never reopened), recovered (Done only after a Failed/Partial/Blocked detour), false-done caught (marked Done, then transitioned back out of Done), or unresolved (no Done yet). The trustworthy-Done rate is the evidence-backed share of all Dones.
• Human interventions, decision vs chore. A counted human touch is either a decision the loop legitimately owed you (a deliberate defer, or a seeded question routed to you) or a chore the self-healing loop should have absorbed (un-parking a revived task, recovering an abandoned run). The chore share, per intent, is the number the autonomy goal drives toward zero.
• Unnecessary loop stops. Every halt into NeedsUser is counted; the ones that were approval or pause friction rather than a real seeded question are reported as reducible waste.

Every number traces to a specific transition or run, never a hand-tally. yardlet trust --json emits the autonomy metrics (nested under done_trust, human_touches, loop_stops, and sources) as machine-readable JSON. The terminal UI shows the same numbers in a Trust panel (the T key on the Home screen).

The Trust Report is read-only across both layers. It reports, but it does not edit .agents/workers.yaml, rewrite routing policy, mark tasks done, or promote harness assets.

Outcome Mining

Outcome Mining uses the same telemetry, but looks for threshold-crossing patterns that are better treated as harness work than as one-off failures. yardlet harness review surfaces mined observations next to learned rules and skills.

Current v0.8 thresholds are:

Signal	Threshold	Why it matters
Worker no-result hotspot	at least 6 runs for the worker, with no-result rate at least 10%	The worker often finished without a parseable result, which points to a packet or output-contract problem.
High-retry task kind	at least 3 tasks of the kind reached Done, averaging at least 2.5 attempts	The kind rarely lands first-pass, which suggests a skill or sharper acceptance criteria may help.

The first signal is per worker. The second is per task kind and only averages tasks that eventually reached Done, so it measures retry effort rather than unresolved failure.

Outcome Mining is suggestions-only. It can name the recurring deterministic outcome and suggest checking a worker contract, writing a rule, creating a skill, or sharpening acceptance criteria. It never changes routing, never edits policy, and never promotes a rule or skill without the normal harness path. Telemetry observes; humans decide which suggestions become durable harness assets.

Source Trail

The public feature claims above are grounded in the v0.8 entries in CHANGELOG.md and the build record in docs/v0.8-decisions.md. The policy boundary matches the existing routing and telemetry design: telemetry suggests, but it never binds runtime policy.