All five items shipped together in v1.29.0 (2026-08-14).
Lifecycle note: like E06, this epic ran planned → done inside one maintainer-directed session
(“lets plan this implementation from 1 to point 5”, then “yes please start implementing”). That
direction is what satisfies the T2 / maintainer-authored requirement — human-directed, not
loop-initiated. The WIP cap was never consumed.
Shared verification: repo-wide anchor check across all tracked markdown, run after BL-0026’s
section rename — zero broken. Every touched doc confirmed under the 1,050-line soft cap
(04 1,020 · 06 358 · 07 544 · 08 671 · 10 754). Markdown-only diff.
| Field | Value |
|---|---|
| Epic | E07-agentic-workflow-pass |
| Pillar | P9 |
| Priority | P1 |
| Effort | S |
| Status | done |
| Test | pass — convention documented with protocol and stated limits; both templates carry the rule |
| Deps | — |
| Lock | — |
Why / Description: The two-layer memory design rests on an assumption nothing checks — that the instruction file is still in context. It is read at session start; as context fills, the earliest content is dropped first, and nothing announces it. The agent keeps working, equally confident, with the project’s rules no longer applying. Adopted a canary: the instruction file requires a short marker at the start of every response, and the marker’s absence is the signal.
Done means:
08 documents the convention, the marker requirement, and — the part that makes it useful —
the protocol when it vanishes (stop, re-read, rehydrate from active context, or restart).Files (actual): methodology/08_lessons_and_memory.md (“The context-integrity canary” plus a
contents-table row); templates/CLAUDE.md, templates/AGENTS.md (“Context integrity” section).
Resolution: Shipped in v1.29.0. Deliberately framed as a smoke alarm, not proof of safety — no gate depends on it, and the section says so explicitly, because a detector that produces both false positives and false negatives is exactly the kind of thing a reader over-trusts. Text marker rather than an emoji, keeping the corpus free of pictographs (a cross-AI review in v1.25.0 flagged our only emoji as an outlier and it was removed).
| Field | Value |
|---|---|
| Epic | E07-agentic-workflow-pass |
| Pillar | P9 |
| Priority | P1 |
| Effort | M |
| Status | done |
| Test | pass — third mode documented; 4 inbound anchors repointed; repo-wide anchor check zero broken |
| Deps | — |
| Lock | — |
Why / Description: Cross-AI validation ran only after work existed — findings-verification on a finished item, diff-verification on a proposed patch. Nothing checked whether the plan itself was factually grounded before implementation. Added spec-verification: a fresh session checks the item body and its Code Map against the real codebase — grounded (do these files and functions exist), coherent (does the approach fit what is actually there), sufficient (would executing it satisfy the goal) — before any code is written.
Done means:
10’s “Two modes” becomes three, with spec-verification first and the same PASS/FAIL plus
citations output shape.Files (actual): methodology/10_testing_and_verification.md (section renamed plus new mode);
methodology/04_backlog_items.md (Code Map points at it); anchors repointed in CHEATSHEET.md,
methodology/07, methodology/09, templates/AUTONOMOUS_LOOP.md.
Resolution: Shipped in v1.29.0. The riskiest item of the five, entirely because of the rename:
#two-modes-... was referenced from four files, and the visible link text was stale in two of them
even after the anchors were repointed — caught by checking text as well as targets. This is the
shift-left of the v1.25.0 routing work: v1.25.0 taught us to route plan-layer defects well; this
moves the check to before the code exists, where the fix is editing a paragraph.
| Field | Value |
|---|---|
| Epic | E07-agentic-workflow-pass |
| Pillar | P1 |
| Priority | P2 |
| Effort | S |
| Status | done |
| Test | pass — rule stated with a one-question test and cross-linked to the memory admission test |
| Deps | — |
| Lock | — |
Why / Description: v1.26.0 added the memory admission test — derivable from source is never stored — but stated it only for memory entries. System documentation had no equivalent, so it kept accumulating implementation detail that goes false the first time someone rewrites the implementation, silently, while still being read as authority.
Done means:
07’s living-documents playbook says document interactions and contracts, not
implementations.08’s admission test as the same principle on a different artifact.Files (actual): methodology/07_definition_of_done.md (“Write docs at the altitude that survives
change”).
Resolution: Shipped in v1.29.0. The framing that earns it: a confidently wrong document costs more than an absent one — the absent one sends you to the code, and the code is never out of date.
| Field | Value |
|---|---|
| Epic | E07-agentic-workflow-pass |
| Pillar | P9 |
| Priority | P2 |
| Effort | XS |
| Status | done |
| Test | pass — routing extended one level up; anti-pattern row added |
| Deps | BL-0026 |
| Lock | — |
Why / Description: Failure-layer routing stopped at the item: intent, plan, code. But a finding can be bigger than any item — an external requirement changed and the architecture no longer fits it. Without naming that layer, the default is to wedge the new requirement into the old architecture, producing a codebase that “works” while every future item pays interest.
Done means:
07’s routing section extends the same rule one level up, and says the fix is a
strategy/pillar-level re-evaluation rather than something an item absorbs quietly.06 gains the matching anti-pattern row.Files (actual): methodology/07_definition_of_done.md (“When the architecture is what’s wrong”);
methodology/06_working_principles.md (anti-pattern table).
Resolution: Shipped in v1.29.0. Smallest item in the epic — it is the architecture-scale sibling of never patch code to compensate for a wrong plan, and reuses that section rather than starting a new one.
| Field | Value |
|---|---|
| Epic | E07-agentic-workflow-pass |
| Pillar | P1 |
| Priority | P2 |
| Effort | S |
| Status | done |
| Test | pass — setting documented in 08 with the cut-filler/keep-reasoning boundary; both templates carry it |
| Deps | — |
| Lock | — |
Why / Description: The corpus had zero guidance on response verbosity, so every contributor either re-negotiated it per session or accepted the default by inertia. It costs twice: tokens, and — larger — reader attention, since a four-paragraph answer to a one-line question trains the reader to skim, and skimming is how a real caveat gets missed.
Done means:
08 names it as an instruction-file setting with a stated default (technical and direct; no
praise preamble, no restating the question).Files (actual): methodology/08_lessons_and_memory.md (“Setting the house verbosity” plus a
contents-table row); templates/CLAUDE.md, templates/AGENTS.md.
Resolution: Shipped in v1.29.0, and deliberately not adopted as the source’s “caveman mode”. Dropping “great question” costs nothing; dropping the why dismantles the methodology — Principle 1 requires stated assumptions, plan-mode requires the reasoning, the DoD requires an honest account of what failed. The section says that an agent which responds to terseness guidance by becoming confidently silent about its uncertainty is worse than a wordy one, and that this means the setting is too aggressive.