Three defects the first sweep left behind, each caught by a rule the guard
did not yet have.
Statuses. RULES.md gives three; the tree had 25 values, and the free-text ones
were carrying progress. "Implemented (wiring landed)" and "Ratified (rev. 2e).
Rev. 2e adds the OpenTaskByPR lookup. Rev. 2d revised section 10..." were both
changelogs living in the one field the other rules exempt. Every PRD now
declares Draft, Approved or Finalised and nothing else; five that declared
nothing now declare one. A revision number moves to **Revision:**, a pointer
to the authoritative document to **Authority:**, and what a past revision
changed to git. CLAUDE.md's citation of state-store-contract rev. 2e still
resolves.
Stale cross-references. "Supersedes the 2026-09-03 open question", "this
bullet previously said", "an earlier draft of this document claimed" all point
at text the reader cannot see, because it was replaced. A supersede marker
survives only when it names a file, a section or a symbol that can be found.
Temporal hedges. 100 uses of today, currently and right now across 31 files.
Each dates its sentence to the day it was written: "[notify] today is one
webhook URL, read in one place" stops being true the moment this design lands,
while still reading as the premise it rests on.
The header-field exemption was also too wide. "**Date:** 2026-09-03, Channel
resolved 2026-09-22" smuggled a dated status into the one line the rules did
not read; the exemption is now a Date holding only a date and a Status holding
only one of the three.
183 findings across 28 of 46 PRDs, of which 150 were file:line citations. Two
were already hard-broken: one named a file that no longer exists, one pointed
past the end of its file. The rest resolved to a line that still exists,
which is the worse failure -- a citation that silently drifted onto unrelated
code still reads as authoritative.
Citations now name the file and the symbol. Dated status markers become the
decision they carried ("Status 2026-09-05: resolved" -> "Resolved"), since
whether a question is answered is design information and when it was answered
is not. Progress inventories become scope statements: mcp-tool-completion now
fences what is out of scope rather than reporting what shipped.
status-health-surface drops phase-completion from its headings and keeps the
rule that earned them, that a reported field must have a truthful source.
eda-playbook-engine.md was the worst file and my own t2db.23 status block was
part of why; it is rewritten as the resolution it records.
Closes archie-core-h019.7. Three parts:
- Promote docs/prds/adversarial-self-review.md into
docs/architecture/adversarial-review.md, describing the stage as shipped
(subagent executor, .git-free snapshot, ReviewReport contract, fail-closed
park, models.reviewer budget, operator-triggered review_pr) and mark the PRD
as promoted.
- Add the review stage to ARCHITECTURE.md's task lifecycle.
- Reconcile CLAUDE.md's Development Protocol: distinguish the contributor's
opt-in manual pass from archied's automatic pre-PR review.
Note: the h019.7 text said the review diverts to waiting_human, but the
shipped stage (h019.1/h019.4) parks to StatusParked, not waiting_human. This
change documents the shipped behaviour (parked), not the stale wording.
Section 1 of the adversarial self-review decision was superseded on
2026-08-24: it assumed daemon-owned in-process execution and a daemon-side
second Pool.Acquire, both removed by archie-core-q15y. It is now re-decided
rather than merely marked dead, so the document is Active again.
The reviewer is an agent.Subagent run from inside the task-scoped worker.
ai-sdk v0.1.22 is already pinned, so this needs no SDK bump, and it adds no
host executor, no second container acquire and no nested container lifecycle
-- the open question the supersession left.
Subagent.Run passes Prompt and never populates Messages, so conversation
isolation is a construction fact rather than prompt text. It does NOT give
filesystem isolation: the sub-agent shares the parent's process, working
directory and credentials, so the .git-stripped snapshot survives from the
original design and is load-bearing, and the reviewer's toolset must be
read-only, rooted at the snapshot, and must exclude the worktreerpc
publication grant -- otherwise a reviewer could push the branch it exists to
block. Run() not Tool(), so the implementer cannot elect to review itself.
Findings return through a closure-captured tool because Run yields only text,
and are read even when Run errors, since captured findings survive.
Also decides fail-closed: a review that fails to RUN parks the task rather
than opening an unreviewed PR. ReviewEnabled is opt-in, so an operator who
turned the gate on wants the gate. This is why h019.2 models "ran and found
nothing" separately from "did not run" -- collapsing them would make an
outage indistinguishable from a clean review.
h019.3 is unblocked and rewritten with the construction rules. h019.4's
description said the task parks as waiting_human, contradicting section 4's
decision of StatusParked; corrected, and the fail-closed rule added.
Autonomous workflows now have exactly one production boundary: a task-scoped archie-agent receives a full-task core-NATS request and runs the worker-local stage loop. The legacy [agent] section remains decode-only for migration; it no longer selects runtime behavior.
Render completed ToolCallEvents as Telegram-safe streaming blocks
with compact name/status and a short cleaned preview. Summarize
file contents and listings instead of dumping source or quoted
JSON. Collapse repeated legacy turn-budget failures.
Remove max_tokens turn gating and the 200000 turn_budget_chars
default so absent or commented config cannot restore an aggregate
cap. Keep per-result max_result_chars and spill.
Confine find/grep discovery to the active workspace/worktree.
Refs #604
archie-core-7d5u.5 (idempotency and poll/webhook parity): the dedup
mechanism (TaskEnvelope.IdempotencyKey + PublishUnique + JetStream
Nats-Msg-Id dedup) and the shared dispatch predicate were already fully
wired by 7d5u.4. What was missing: proof the dedup actually holds, and
observability for a wedged receiver.
- internal/infrastructure/eventbus/nats/dedup_test.go: regression test
proving PublishUnique suppresses a duplicate publish on the same
WorkQueuePolicy stream ARCHIE_TASKS uses in production.
- internal/forge/webhook/receiver.go: Receiver.Status() tracks
authenticated deliveries separately from successful publishes; a GET
to the receiver's listen address (GitHub only ever POSTs) now returns
it as JSON, so a wedged/non-delivering receiver is observable instead
of indistinguishable from an idle repo.
Poll/webhook parity (disabling webhook still polls) needed no code
change: Cycle()/Run() never reference cfg.Forge.Intake anywhere in
internal/daemon, only setupForgeWebhook gates the webhook HTTP server.
archie-core-h019.1 (design: adversarial self-review stage): one-page
decision at docs/prds/adversarial-self-review.md. Reviewer runs in a
second workspace pointed at a .git-stripped snapshot of the reviewed
commit, never the implementer's worktree, so isolation is a property of
what's mounted rather than a prompt instruction. Findings contract is
workflow.ReviewFinding, owned by internal/domain/workflow. A surviving
error-level finding sets StatusParked, not WaitingHuman -- the epic's
informal wording named two different states in the actual machine; this
resolves which.
Both changes reviewed by a fresh subagent with no memory of the
implementation, per CLAUDE.md's adversarial-verification requirement;
zero blocking findings, two minor coverage gaps closed.