project-loop-harness 0.2.3__tar.gz → 0.3.0__tar.gz

{project_loop_harness-0.2.3/src/project_loop_harness.egg-info → project_loop_harness-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: project-loop-harness
-Version: 0.2.3
+Version: 0.3.0
 Summary: Local control plane for agentic project development loops
 Author: Project Loop Harness Contributors
 License: MIT
@@ -26,6 +26,8 @@ License-File: LICENSE
 Provides-Extra: dev
 Requires-Dist: pytest>=8; extra == "dev"
 Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: setuptools>=69; extra == "dev"
+Requires-Dist: wheel; extra == "dev"
 Dynamic: license-file
 
 # Project Loop Harness

project_loop_harness-0.3.0/agent-tasks/0102-source-drift-health-warning.md

@@ -0,0 +1,62 @@
+# 0102: Source drift must surface as evidence health warning
+
+Milestone: v0.2.4 Trust Patch
+Priority: P1
+Area: evidence
+Origin: docs/project-loop-harness-v0.2.3-third-party-review.md P1-1 (verified against main)
+
+## Problem
+
+`_collect_copied_member_findings` in `src/pcl/evidence.py` emits `source_drifted`
+findings (detail `missing` or `size_mismatch`) when the original source of a
+copied evidence member has drifted. But `source_drifted` is not in
+`ADHOC_WARNING_FINDING_CODES` (evidence.py:36), so `_adhoc_assessment` returns
+`health: "ok"` when the only findings are `source_drifted`. Agents and humans
+read "ok" and miss the provenance/freshness problem.
+
+## Scope
+
+1. Add `"source_drifted"` to `ADHOC_WARNING_FINDING_CODES` in
+   `src/pcl/evidence.py`.
+2. Confirm the rendering path in `src/pcl/validators.py` (the
+   `elif code == "source_drifted":` branch around line 977) stays consistent
+   with the new classification — the finding text must keep re-rendering
+   identically through the shared `assess_adhoc_evidence` path (0095 pattern).
+3. Tests that pin the semantics:
+   - copied artifact intact + source file deleted → `health: "warning"` with a
+     `source_drifted` finding, `detail: "missing"`.
+   - copied artifact intact + source size mismatch → `health: "warning"` with
+     `detail: "size_mismatch"`.
+   - copied artifact intact + source intact → `health: "ok"` (no regression).
+   - copied artifact corrupted (`copy_hash_mismatch`) + source drifted →
+     still `warning` (both findings present, no double-count crash).
+4. `pcl validate --strict --json` must treat `source_drifted` as a warning that
+   remains a warning under strict (extend the existing
+   `_strict_warning_remains_warning` coverage), not an error, and must stay
+   green overall.
+
+## Invariants (what to protect, on the normal paths)
+
+- The copied artifact stays usable: no finding may remove or rewrite copied
+  files under `.project-loop/evidence/adhoc-files/`; assessment is read-only.
+- The findings array shape is additive-only: existing finding codes, their
+  `path`/`detail` fields, and their ordering semantics are unchanged.
+- Health values remain the existing three-level `ok | warning | error`; do not
+  introduce `artifact_health` / `source_health` split fields (explicitly
+  deferred by approval 2026-07-08).
+
+## Non-scope
+
+- artifact_health / source_health separation (deferred until dogfood
+  source_drift_rate data exists).
+- Schema or migration changes.
+- Context pack / dashboard surfacing changes beyond what the shared
+  assessment already propagates.
+
+## Acceptance
+
+- All tests above green; full `pytest` green.
+- Live smoke: create copied evidence in a scratch project, delete the source,
+  run `pcl validate --strict --json` → evidence health shows `warning`,
+  strict stays `ok: true`.
+- Release note entry describing the semantic change.

project_loop_harness-0.3.0/agent-tasks/0103-security-policy-v02x.md

@@ -0,0 +1,46 @@
+# 0103: SECURITY.md v0.2.x update and copied-evidence policy
+
+Milestone: v0.2.4 Trust Patch
+Priority: P1
+Area: docs/security
+Origin: docs/project-loop-harness-v0.2.3-third-party-review.md P1-2 (verified: SECURITY.md still says `0.1.x` while the released line is 0.2.x)
+
+## Problem
+
+SECURITY.md states the current public release line is `0.1.x`, contradicting
+the v0.2.3 release. Additionally, v0.2.3 introduced `pcl evidence add --copy`,
+which retains copies of source files under
+`.project-loop/evidence/adhoc-files/` — a new secret-retention surface that the
+security policy does not mention.
+
+## Scope
+
+Update `SECURITY.md` only (no code changes):
+
+1. Supported versions table: `0.2.x` = supported, `<0.2` = not supported.
+2. Add a "Copied evidence" section stating:
+   - `.project-loop/evidence/adhoc-files/` may contain copied sensitive source
+     files when `--copy` is used.
+   - Copied evidence must not be committed unless intentionally curated;
+     `.project-loop/` is gitignored by default and should stay so.
+   - Redaction is caller responsibility; `pcl` performs path-shape sensitive
+     guards (0096) but is not a secret scanner (established v0.1.11 decision —
+     keep this wording).
+   - MCP / read-only exposure must not reveal raw evidence contents by
+     default.
+   - Dashboard HTML is a human view, not a machine context source.
+3. Note that the release checklist (0106) will include a SECURITY.md
+   supported-versions check per release.
+
+## Invariants
+
+- Do not weaken any existing statement in SECURITY.md; changes are additive or
+  version-string corrections.
+- Keep the established epistemic vocabulary (claims-not-facts; no "scanned /
+  guaranteed secret-free" style assertions).
+
+## Acceptance
+
+- SECURITY.md shows `0.2.x` as the supported line.
+- Copied evidence risks and commit policy are stated.
+- `pytest` green (docs-only change; no test changes expected).

project_loop_harness-0.3.0/agent-tasks/0104-python-ci-matrix.md

@@ -0,0 +1,44 @@
+# 0104: Python 3.10–3.13 CI matrix
+
+Milestone: v0.2.4 Trust Patch
+Priority: P2
+Area: ci
+Origin: docs/project-loop-harness-v0.2.3-third-party-review.md P2-1 (verified: ci.yml pins `python-version: '3.11'` while pyproject classifiers list 3.10/3.11/3.12/3.13)
+
+## Problem
+
+`pyproject.toml` declares `requires-python = ">=3.10"` and classifiers for
+3.10–3.13, but CI runs a single Python 3.11. The advertised support range is
+not continuously verified.
+
+## Scope
+
+Update `.github/workflows/ci.yml`:
+
+1. Run lint + tests + smoke on a matrix of `["3.10", "3.11", "3.12", "3.13"]`:
+   - `ruff check .`
+   - `pytest`
+   - smoke: `pcl --version`; then in a fresh temp directory `pcl init`
+     (note: `pcl init` takes no positional argument, run it inside the temp
+     cwd), `pcl validate --strict --json`, `pcl render --json`.
+2. Keep build / twine check / sdist contract steps on a single canonical
+   version (3.12).
+3. Keep total CI time reasonable — matrix jobs run in parallel; do not add
+   new workflow files unless the existing one cannot express it.
+
+## Invariants
+
+- Existing CI steps (including the advisory retrieval eval step from 0080)
+  keep their current blocking/advisory semantics; the matrix must not turn an
+  advisory step into a blocking one or vice versa.
+- No changes to pyproject metadata, package code, or tests are expected. If a
+  Python-version-specific test failure surfaces, report it in the task output
+  rather than papering over it with skips — that finding is the point of this
+  task.
+
+## Acceptance
+
+- CI green on all four Python versions (verify with a branch push / PR run or
+  `gh run watch` after pushing the branch).
+- Classifiers and CI matrix agree.
+- Release note mentions the matrix result.

project_loop_harness-0.3.0/agent-tasks/0105-evidence-copy-observability.md

@@ -0,0 +1,51 @@
+# 0105: Evidence copy observability
+
+Milestone: v0.2.4 Trust Patch
+Priority: P2
+Area: evidence
+Origin: docs/project-loop-harness-v0.2.3-third-party-review.md P2-2
+Depends on: 0102 (both touch src/pcl/evidence.py — start after 0102 merges)
+
+## Problem
+
+v0.2.3 serialized evidence ID allocation to fix a concurrency race. When
+`evidence add --copy` copies large files, the copy work may extend the time a
+SQLite write transaction/lock is held. We have no measurements, so we cannot
+judge whether reserved-row or counter designs are ever needed. Observability
+first, redesign only if data demands it.
+
+## Scope
+
+1. Measure and record, for `pcl evidence add --copy` invocations:
+   - `copy_duration_ms` (wall time of the file-copy phase)
+   - `copied_total_bytes`
+   - `member_count`
+   These go into the `evidence_added` event payload (additive fields) via the
+   standard `append_event` path (both stores, per the 0089 lesson — no
+   JSONL-only or SQLite-only side channels).
+2. Where measurable without redesign, record lock-related diagnostics (e.g.
+   time spent inside the write transaction) as an additive payload field; if
+   it is not cleanly measurable, state so in the task output instead of
+   adding invasive instrumentation.
+3. Add a concurrent copy stress test: N parallel `evidence add --copy`
+   processes against one project must all succeed deterministically with
+   unique IDs (extend the existing 0101 race-test pattern).
+
+## Invariants
+
+- Timestamps/durations must not break determinism guarantees of existing
+  contracts: duration fields live only in event payloads, not in receipts,
+  context packs, manifests, or anything byte-compared by tests.
+- ID allocation semantics are unchanged (serialized allocation from 0101 is
+  kept as-is).
+- Non-`--copy` `evidence add` behavior and payload shape are unchanged
+  except for fields explicitly listed above being absent.
+- No new tables, no migration.
+
+## Acceptance
+
+- `evidence add --copy --json` emits an event whose payload includes the three
+  metrics; existing consumers/tests unaffected (additive-only).
+- Concurrent stress test green and deterministic across repeated runs.
+- Full `pytest` green; `pcl validate --strict --json` green after live copy
+  operations in a scratch project.

project_loop_harness-0.3.0/agent-tasks/0106-release-checklist.md

@@ -0,0 +1,21 @@
+# 0106: Release checklist contract
+
+Milestone: v0.2.4 Trust Patch
+Priority: P2
+Area: docs/release
+Origin: docs/project-loop-harness-v0.2.3-third-party-review.md PLH-0108
+Implementation: orchestrator-authored — deliverable is `docs/release-checklist.md`
+
+## Goal
+
+Make every release reproducible by contract, capturing the steps and traps
+already proven across v0.1.7–v0.2.3 releases.
+
+## Acceptance
+
+- `docs/release-checklist.md` exists and covers: version bump, release note,
+  ruff, pytest (matrix), editable-install sanity, build, twine check, sdist
+  contract verification, fresh-venv wheel smoke, `pcl validate --strict
+  --json`, `pcl render --json`, SECURITY.md supported-versions check, tag →
+  GitHub Release → trusted publishing → PyPI verification, local pipx update.
+- The v0.2.4 release is executed against this checklist.

project_loop_harness-0.3.0/agent-tasks/0107-agent-tasks-backlog-index.md

@@ -0,0 +1,20 @@
+# 0107: agent-tasks backlog index
+
+Milestone: v0.2.4 Trust Patch
+Priority: P3
+Area: docs/backlog
+Origin: docs/project-loop-harness-v0.2.3-third-party-review.md PLH-0109
+Implementation: orchestrator-authored — deliverable is `agent-tasks/README.md`
+
+## Goal
+
+Make the `agent-tasks/` directory navigable: distinguish active backlog from
+completed design history without duplicating `TASKS.md` (which stays the
+canonical ordered index).
+
+## Acceptance
+
+- `agent-tasks/README.md` exists: explains the spec-first workflow, points to
+  `TASKS.md` for the full history, and lists the active backlog with
+  milestone/priority/status.
+- New contributors and agents can find the next task from it.

project_loop_harness-0.3.0/agent-tasks/0108-target-bound-code-context-receipts.md

@@ -0,0 +1,152 @@
+# 0108: Target-bound code context receipts
+
+Milestone: v0.3.0 Target-Bound Context
+Priority: P1
+Area: code-context/context-pack
+Origin: docs/growth-plan-v0.2.4-v0.5.md v0.3.0; docs/project-loop-harness-v0.2.3-third-party-review.md P2-3
+Depends on: 0113 (evidence_links table, migration 007) — merge first.
+
+> **Revision 2026-07-08 (坂本承認):** the original no-migration invariant is
+> **retracted**. Target binding is now persisted as a queryable `evidence_links`
+> row (`link_role="code_context"`, from 0113) **and** embedded in the receipt
+> artifact; receipt selection is a SQL query, not an artifact scan. This unblocks
+> `pcl context check` (v0.3.1) and v0.4 `bound_receipt_coverage` without a second
+> migration later.
+
+## Problem
+
+`pcl context pack --include-code-context` currently reads the latest
+`context_receipt` evidence row. The pack tells the reader that this is
+`scope: "unscoped_latest"` with `binding_strength: "none"`, which is honest,
+but it still risks handing a worker a receipt from an unrelated task or job.
+
+The existing docs reserve `scope: "target_bound"` and
+`binding_strength: "caller_asserted"`, but there is no command path that
+creates or requires such a receipt. v0.3.0 turns that reserved contract into a
+narrow, explicit handoff workflow, backed by queryable link state.
+
+## Scope
+
+1. Extend `pcl impact --diff` with explicit target flags:
+   - `--for-task T-XXXX`
+   - `--for-job J-XXXX`
+   Exactly zero or one target flag is allowed. Existing diff mode flags
+   (`--base`, `--staged`, `--unstaged`, `--include-untracked`,
+   `--all-changes`, and provided diff input) keep their current semantics.
+2. Validate the named target before writing a receipt. Missing or malformed
+   target IDs must fail with a typed error and must not create evidence or an
+   `evidence_links` row.
+3. When a target flag is present, record the binding in **two** places, in the
+   same transaction that writes the receipt evidence:
+   - An `evidence_links` row (0113 helper): `(evidence_id=<receipt>,
+     target_type, target_id, link_role="code_context", created_at)`, where
+     `target_type` is `task` or `agent_job`.
+   - Target binding metadata inside the `impact/v0` response and the
+     `context-receipt/v0` artifact so the receipt is self-describing:
+     ```json
+     {
+       "target_binding": {
+         "target_type": "task",
+         "target_id": "T-0001",
+         "binding_strength": "caller_asserted",
+         "source": "impact_flag"
+       }
+     }
+     ```
+     For job targets, use `target_type: "agent_job"`. The link row and the
+     artifact `target_binding` must agree.
+4. Update context-pack receipt selection when `--include-code-context` is used.
+   Selection is a **query on `evidence_links`** (0113 helper), not an artifact
+   scan:
+   - For a task pack, prefer the newest receipt evidence with a matching
+     `evidence_links` row `(target_type="task", target_id=<task>,
+     link_role="code_context")`.
+   - For a job pack, prefer the newest matching
+     `(target_type="agent_job", target_id=<job>, link_role="code_context")`.
+   - If no matching bound receipt exists, preserve the current unscoped latest
+     fallback unless the caller requires a bound receipt.
+5. Add `pcl context pack --require-bound-receipt`, valid only with
+   `--include-code-context`. When no matching bound receipt exists, fail with a
+   typed error `context_pack_bound_receipt_required` and include a
+   target-specific refresh suggestion:
+   - `pcl impact --diff --for-task T-XXXX --json`
+   - `pcl impact --diff --for-job J-XXXX --json`
+6. Update `code_context.relevance` in the embedded summary:
+   - Matching bound receipt: `scope: "target_bound"`,
+     `binding_strength: "caller_asserted"`, plus `target_type` and
+     `target_id`.
+   - Unbound fallback: keep `scope: "unscoped_latest"` and
+     `binding_strength: "none"` with an explicit warning.
+   - Missing receipt: keep `scope: "missing_receipt"` and
+     `binding_strength: "none"`.
+7. Update docs and CLI help for `docs/code-context.md`,
+   `docs/context-pack.md`, and relevant tests.
+
+## Invariants (what to protect, on the normal paths)
+
+- Binding is recorded in lockstep: when a target flag is given, the receipt
+  evidence row, its `evidence_links` `code_context` row, and the artifact
+  `target_binding` are written in one transaction and must agree. Never write
+  the link row without the receipt, or an artifact `target_binding` without the
+  link row.
+- Binding is a **caller assertion**, not PLH proof that the receipt is
+  semantically relevant. Use claims-not-facts vocabulary; do not add fields such
+  as `safe_to_continue`, `verified_relevant`, or `agent_read`.
+- Extend `impact/v0`, `context-receipt/v0`, and `context-pack/v1` **additively**
+  via the `code-context-summary/v0` insulation layer (0078). The pack embeds the
+  summary and `receipt_ref` (evidence_id / receipt_path) only — never the full
+  receipt body or copied evidence contents.
+- `pcl context pack` remains read-only: it must not run `pcl index build`,
+  `pcl impact`, Git commands, tests, or any artifact-generating command. It
+  reads `evidence_links` and evidence rows only.
+- `source_commands` remain read-only re-fetch commands. Artifact-generating
+  commands (`pcl impact …`) appear only under `suggested_refresh_commands`.
+- Existing unscoped latest behavior remains available when
+  `--require-bound-receipt` is absent.
+- Deterministic generated artifacts stay deterministic; timestamped facts remain
+  in receipt/evidence payloads where timestamps already exist.
+
+## Non-scope
+
+- Semantic or embedding-based relevance proof.
+- Automatic target inference from changed files, branches, or task text.
+- `pcl context pack --receipt ...` manual receipt selection.
+- `pcl receipt list` / `pcl evidence link` CLI verbs.
+- Dashboard redesign or MCP surface changes.
+- Hosted code analysis, cloud sync, telemetry, or external LLM calls.
+- `pcl context check` preflight (deferred to v0.3.1).
+
+## Acceptance
+
+- `pcl impact --diff --for-task T-0001 --json` writes a receipt with
+  `target_binding.target_type == "task"` and `target_id == "T-0001"`, and an
+  `evidence_links` row `(target_type="task", target_id="T-0001",
+  link_role="code_context")`.
+- `pcl impact --diff --for-job J-0001 --json` writes a receipt with
+  `target_binding.target_type == "agent_job"` and `target_id == "J-0001"`, and
+  the matching `code_context` link row.
+- Invalid target IDs and mutually exclusive target flags fail with typed errors
+  before any evidence or link row is created.
+- `pcl context pack --task T-0001 --include-code-context --json` chooses the
+  newest matching task-bound receipt over a newer unbound receipt, via the
+  `evidence_links` query.
+- `pcl context pack --job J-0001 --include-code-context --json` chooses the
+  newest matching job-bound receipt over a newer unbound receipt.
+- Without a matching bound receipt and without `--require-bound-receipt`,
+  context pack still succeeds with the existing unscoped latest warning and a
+  target-specific suggested refresh command.
+- With `--require-bound-receipt` and no matching bound receipt, context pack
+  fails with `context_pack_bound_receipt_required` and does not silently fall
+  back to unscoped latest.
+- `source_commands` for code-context packs still contain only read-only
+  commands; `pcl impact` appears only under `suggested_refresh_commands`.
+- Full `pytest` green.
+- Live smoke in a scratch project:
+  1. `pcl init`
+  2. create a goal and task
+  3. `pcl index build --json`
+  4. make a small tracked change
+  5. `pcl impact --diff --for-task T-0001 --json`
+  6. `pcl context pack --task T-0001 --include-code-context --require-bound-receipt --json`
+  7. `pcl validate --strict --json`
+  8. `pcl render --json`

project_loop_harness-0.3.0/agent-tasks/0113-evidence-links-table.md

@@ -0,0 +1,118 @@
+# 0113: Generic evidence links table (migration 007)
+
+Milestone: v0.3.0 Target-Bound Context
+Priority: P1
+Area: schema/evidence
+Origin: docs/growth-plan-v0.2.4-v0.5.md v0.3.0; third-party review 議題1 — migration
+approved and 0108 no-migration invariant retracted (Sakamoto approved 2026-07-08).
+
+## Problem
+
+Evidence-to-target linkage is currently a single `evidence.linked_task_id`
+column (migration 006), which can only express "supporting evidence for a
+task". v0.3.0 introduces a second link role — a code-context receipt bound
+to a task or job (0108) — and v0.3.2 will add master-trace / intent-index
+roles. A single typed column cannot express a role, and adding one column per
+target type (`linked_job_id`, `linked_goal_id`, and so on) sprawls.
+
+Introduce a generic, queryable `evidence_links` table so target-bound receipt
+selection (0108), `pcl context check` (v0.3.1), and v0.4 coverage KPIs become
+SQL queries rather than artifact scans. This is a control-plane primitive, not
+a per-feature column.
+
+## Scope
+
+1. Add migration `007_evidence_links.sql`:
+   ```sql
+   PRAGMA foreign_keys = ON;
+
+   CREATE TABLE IF NOT EXISTS evidence_links (
+     evidence_id TEXT NOT NULL REFERENCES evidence(id),
+     target_type TEXT NOT NULL,
+     target_id   TEXT NOT NULL,
+     link_role   TEXT NOT NULL,
+     created_at  TEXT NOT NULL,
+     PRIMARY KEY (evidence_id, target_type, target_id, link_role)
+   );
+
+   CREATE INDEX IF NOT EXISTS idx_evidence_links_target
+     ON evidence_links(target_type, target_id, link_role, created_at);
+
+   -- Backfill existing task-linked evidence as role 'supporting'.
+   INSERT OR IGNORE INTO evidence_links(evidence_id, target_type, target_id, link_role, created_at)
+   SELECT id, 'task', linked_task_id, 'supporting', created_at
+   FROM evidence
+   WHERE linked_task_id IS NOT NULL;
+   ```
+   Vocabulary in use now: `target_type` is `task` or `agent_job`;
+   `link_role` is `supporting` or `code_context`. Reserve (do not emit yet)
+   `goal` / `workflow_run` / `decision` / `escalation` and
+   `master_trace` / `intent_index` / `worker_output`.
+2. Write path: `pcl evidence add --task T-XXXX` (evidence.py) keeps setting the
+   `evidence.linked_task_id` column (0.x compat) and also inserts an
+   `evidence_links` row `(evidence_id, 'task', task_id, 'supporting', now)` in
+   the same transaction, before the existing `append_event`.
+3. Read path: `context.py._linked_task_evidence` selects from `evidence_links`
+   where `target_type='task' AND link_role='supporting'`, joined to `evidence`
+   for the display columns, ordered `created_at, evidence_id`. When the
+   `evidence_links` table is absent (pre-migration DB opened by a newer binary
+   mid-upgrade), fall back to the current `linked_task_id` column query. Output
+   shape, ordering, and member/stored-path resolution are unchanged.
+4. Strict validation (`validators.py`): add a dangling-link check modeled on the
+   verification `receipt_evidence_id` check (validators.py:596) — every
+   `evidence_links.evidence_id` must resolve to an existing evidence row (the
+   existing `PRAGMA foreign_key_check` already covers this; keep an explicit,
+   message-consistent check too), and for known target types the `target_id`
+   must exist (`task` maps to `tasks`, `agent_job` maps to `agent_jobs`).
+   Unknown `target_type` values are tolerated with no error (forward
+   compatibility).
+5. Provide two small internal helpers for 0108 to reuse (no new CLI surface):
+   insert a link row, and select the newest `evidence_id` for a given
+   `(target_type, target_id, link_role)`.
+
+## Invariants (what to protect, on the normal paths)
+
+- The canonical write path for a task-linked ad-hoc evidence is, in one
+  transaction: the `evidence` row plus the `linked_task_id` column plus the
+  `evidence_links` row plus the `adhoc_evidence_recorded` event. Do not drop the
+  `linked_task_id` column write (compat). Do not write an `evidence_links` row
+  without its mirroring `evidence` row.
+- The migration backfill is idempotent (`INSERT OR IGNORE`) and additive: it
+  must not update or delete any `evidence` row, and re-running `pcl migrate`
+  must be a no-op.
+- `evidence_links` is written only through `pcl` (evidence add today, 0108
+  tomorrow). No command deletes links in this task; no raw-SQL mutation path.
+- The existing task context-pack `linked_evidence` output — shape, ordering,
+  member/stored paths — is unchanged for existing DBs after backfill.
+- Additive only: no existing table or column is dropped or repurposed. Schema
+  version advances 6 to 7; `never-downgrade` and DB-ahead typed rejection (0076)
+  continue to hold.
+
+## Non-scope
+
+- `pcl evidence link` / `unlink` CLI verbs (links are written implicitly by
+  `evidence add` and, in 0108, by `impact --for-task/--for-job`).
+- Job / goal / decision linking commands beyond the helper 0108 needs.
+- Dashboard changes. The existing `heading_evidence_links` locale string is an
+  unrelated UI heading, not this table.
+- Removing or deprecating `linked_task_id` (deferred past 0.x).
+- Reading `evidence_links` from the code-context receipt selection path (that is
+  0108).
+
+## Acceptance
+
+- Fresh `pcl init` gives schema v7, `evidence_links` present,
+  `pcl validate --strict --json` green.
+- A v6 to v7 live migration backfills every `linked_task_id` row into
+  `evidence_links` as role `supporting`; backfilled row count equals the number
+  of non-null `linked_task_id` rows; a second `pcl migrate` changes nothing.
+- `pcl evidence add --task T-0001` writes both the column and the
+  `evidence_links` row; the task context pack lists the evidence identically to
+  before.
+- A hand-broken dangling link (an `evidence_links` row whose `target_type='task'`
+  `target_id` has no matching task) is reported by `validate --strict`; a row
+  with an unknown `target_type` is tolerated (no error).
+- An older (v6-aware) `pcl` binary typed-rejects a v7 DB with
+  `schema_version_ahead` (0076 behavior preserved).
+- Full `pytest` green; live smoke in a scratch project (`init`, then `evidence
+  add --task`, then `validate --strict`, then context pack).

project_loop_harness-0.3.0/agent-tasks/0114-source-hash-drift.md

@@ -0,0 +1,78 @@
+# 0114: Source hash drift detection for copied evidence
+
+Milestone: v0.3.0 Target-Bound Context
+Priority: P2
+Area: evidence
+Origin: docs/growth-plan-v0.2.4-v0.5.md 論点2; third-party review P1 / 議題2 —
+default-on approved (坂本承認 2026-07-08). Extends 0102.
+
+## Problem
+
+`_collect_copied_member_findings` in `src/pcl/evidence.py` (the source-freshness
+block at evidence.py:589-605) detects source drift only via `detail: "missing"`
+and `detail: "size_mismatch"`. A **same-size content change** to the source
+file is undetected: `source_drifted` never fires, so a copied evidence member
+whose origin was silently rewritten to a different-but-equal-length body reads
+as fresh. The recorded member `sha256` (`expected_sha256`) is already a
+parameter of this function (evidence.py:545) but is never compared against the
+current source.
+
+0102 already classifies `source_drifted` as an evidence-health **warning**, so
+closing this gap needs no health-model change — only the missing hash
+comparison.
+
+## Scope
+
+1. In `_collect_copied_member_findings`, extend the `else` branch that currently
+   only handles the size check: when the source exists **and**
+   `source_stat.st_size == size_bytes`, compute the source file sha256 (reuse
+   `_sha256_file`, evidence.py:907) and compare to `expected_sha256`.
+   - Mismatch → append `{"code": "source_drifted", "path": member_path,
+     "detail": "hash_mismatch"}`.
+   - `OSError` while reading the source → treat as unreadable and reuse the
+     existing `detail: "missing"` finding (do not crash, do not add a new code).
+2. Default-on, no `--deep` flag. The hash is computed **only** in the same-size
+   branch, so `missing` and `size_mismatch` stay the cheap fast paths. The
+   `copy_max_member_bytes` cap (10MB, 0097) bounds per-member cost.
+3. Confirm the `validators.py` `source_drifted` rendering branch (the
+   `elif code == "source_drifted":` branch, ~line 977) renders the new
+   `hash_mismatch` detail consistently through the shared
+   `assess_adhoc_evidence` path (0095 pattern) — health stays `warning`.
+
+## Invariants (what to protect, on the normal paths)
+
+- Read-only assessment: no finding rewrites or removes copied artifacts under
+  `.project-loop/evidence/adhoc-files/` or any source file.
+- Findings shape is additive-only: `source_drifted` gains one new `detail`
+  value, `hash_mismatch`. Existing `missing` / `size_mismatch` details, their
+  `path` field, and ordering are unchanged.
+- Copied-artifact health stays semantically **separate** from source freshness
+  in wording (0102 invariant): `copy_hash_mismatch` = the copied artifact broke;
+  `source_drifted/hash_mismatch` = the origin changed since the copy. Do not
+  merge or cross-reference the two codes.
+- Health stays the three-level `ok | warning | error`; no `artifact_health` /
+  `source_health` split (still deferred until dogfood `source_drift_rate` data).
+
+## Non-scope
+
+- A `--deep` / lightweight-vs-deep mode split (unnecessary: the size gate plus
+  the 10MB member cap already bound cost).
+- Hashing sources of **reference** (non-copied) members — reference members are
+  already hash-checked against the source directly by `_assess_reference_member`.
+- `artifact_health` / `source_health` field separation.
+
+## Acceptance
+
+- Test: copied artifact intact + source rewritten to different content of the
+  **same byte length** → `health: "warning"` with a `source_drifted` finding,
+  `detail: "hash_mismatch"`.
+- Test: copied artifact intact + source intact → `health: "ok"` (no regression).
+- Test: `copy_hash_mismatch` (artifact corrupted) + `source_drifted/hash_mismatch`
+  co-occur → still `warning`, both findings present, no double-count crash.
+- `pcl validate --strict --json` treats `source_drifted/hash_mismatch` as a
+  warning that remains a warning under strict (extend
+  `_strict_warning_remains_warning` coverage) and stays green overall.
+- Live smoke: in a scratch project, `evidence add --copy` a file, edit the
+  source to same-length different content, `pcl validate --strict --json` →
+  evidence health `warning`, detail `hash_mismatch`, strict `ok: true`.
+- Full `pytest` green; release note entry describing the tightened semantics.