bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle

Re-grilled 2026-10-26 with the maintainer, against origin/develop at 579b4ef4 (plugin 0.31.0).
The ticket is split. This one keeps the verdict half: /build opens no pull request on a red
verdict, ends cancelled, and keeps the work on the remote branch. The grouping-ticket half
(old criteria 7 to 13, 16 and 18, and the grouping part of 17) moved to
feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket, which waits for the Consumer's tracker to be
redefined. Each criterion below carries holds, amended, new or moved. The DAG shape,
the Archon version and #652 are ruled under § Rulings of 2026-10-26.

Amended 2026-09-04 after a grilling with the maintainer. The criteria below replace the
original body's two questions. Class ruling, under #456
§ Notes: this is a second variable — it changes what /build produces and what /tickets
creates, and neither effect separates from the mechanism checks run 3 measures. So #439 has left
#457's blockers and belongs to run 4's change set. The maintainer's words: "ok, espera a run 3".

What this is

Two claims were filed here. Both were measured during the grilling and neither survived as written.

The first claim, that open-pr running before its gate is what let run 2 ship a pull request. In
run 2 gates.build was afk, so build-pr-gate was skipped by its own when condition wherever it
sat in the DAG. Reversing the two nodes would have changed nothing about run 2. The ordering and the
run-2 outcome are unrelated.

The second claim, that a reject at /pr-review's review-gate destroys every finding. It does
not. synthesize writes the findings and the summary sections, and reconcile writes the composed
summary and the thread plan, all four to the Session dir and all four before the gate pauses. The
findings are on disk when a human rejects. What destroys them is archon complete <branch> --force
deleting the worktree, which is a cleanup defect and not a gate defect.

What is left, and it is the real defect: a build that fails its own acceptance criteria opens a
pull request. evidence withholds evidence.json on a red verdict, the engine refuses terminal
completed, and the run ends failed — with a complete, mergeable pull request already open. That
is ADR-0034
§ Decision item 4's deferred trade, reopened: "If that cost is judged too high, gate open-pr on the
verdict; that is a separate decision with its own blast radius."

And a second defect fell out of the measurement. The one artefact the engine demands is the only
one that carries no information. evidence_policy checks presence, never content; and the file,
when it exists, is { slug, verification_failures: [], goals_check_failures: [] } with both arrays
empty by construction, because the node returns before writing whenever either is non-empty. It can
never say what failed.

The decision

evidence always writes evidence.json and reports a verdict. A verdict other than pass stops the
run before any pull request exists. The run ends cancelled, and the work stays on the remote branch
with report.md inside it. The cancel: message tells the operator where to find it.

How a human who is not watching the run learns about a red build is not decided here. That is the
grouping ticket, moved to feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket.

Rulings of 2026-10-26

The DAG shape is the sibling idiom, not a chain. After the node that commits and pushes, two
siblings depend on it. The cancel: node is a leaf with when on a verdict other than pass. The
node that opens the pull request has when on a verdict equal to pass, and no trigger_rule.
guard-not-ready in /build and guard-no-diff-base in /pr-review already use this idiom. The
pass-only comparison sends unresolved to the cancel side without a third branch. It also drops
the trigger_rule: all_done that the 2026-09-04 body called the criterion most easily satisfied
wrongly. Criterion 0 measures that the sibling never runs on red.
The Archon behaviours are measured on 0.12.1 before any Box YAML changes. Run 4 runs on 0.12.1.
The 2026-09-04 probe ran on 0.7.0, so its results are not evidence for this version. Criterion 0
runs the probe again.
#652 lands first. After this ticket a
red branch has no pull request. A /qa run on that branch would then meet bug(unic-archon-dlc): /qa ends green on paths where nothing merged #652: no open pull
request, merge runs, and the run ends green with nothing merged. bug(unic-archon-dlc): /qa ends green on paths where nothing merged #652 blocks this ticket.
Criteria name properties a reader can confirm, never counts. The counts in old criteria 13 and
19 went stale between two readings.

Acceptance criteria

New. Before any Box YAML changes, a probe on Archon 0.12.1 measures the four behaviours the
criteria below rest on. Run it by path, with an isolated ARCHON_HOME, in a scratch repository
outside every clone, as #505's and
#520's criterion 1 did. The four:
1. A script: node that prints a JSON object on stdout can be read field by field in a when
  expression, such as a field verdict compared with 'pass'.
2. A cancel: node whose when holds sets terminal status cancelled, and a sibling with the
  inverse when does not run, nor does any node that depends on that sibling.
3. On the other branch, the skipped cancel: leaf does not stop the run from reaching completed.
4. A cancel: message substitutes node-output references.
Stop rule: if any of the four does not hold, stop and report what you measured. Do not adapt
the design in the implementation. The ADR records the probe's date, the Archon version and each
result.
Holds (checked at 579b4ef4: evidence still calls process.exit(0) before writing on any
verdict other than pass). evidence writes evidence.json on every verdict, and the file
carries the verdict. The node no longer returns early. The file names the verdict it recorded and
lists the failures that produced it, so a reader learns what failed from the file the engine
demands. The Session-dir mirror is written on every verdict too.
Amended (at 579b4ef4, evidence already prints { verdict, test_outcome, unresolved_needs }
on stdout, added by #430, and its comment
says nothing reads it yet). The verdict evidence prints on stdout is what the two siblings'
when expressions read. Stdout stays JSON only, and every human line stays on stderr. The comment
that says nothing reads it yet is rewritten to name its readers.
Holds (the enum pass | fail | unresolved exists at 579b4ef4). unresolved is not green.
Both when expressions compare with 'pass' only, so unresolved and fail take the same branch.
No expression names fail or unresolved.
Holds. A verdict other than pass opens no pull request, and the run ends cancelled. A
cancel: node fires on that verdict and sets terminal status cancelled, distinct from failed.
Amended (the chain with trigger_rule: all_done is replaced by the sibling idiom, see § Rulings).
The node that opens the pull request and the cancel: node are siblings. Both depend on the
node that commits and pushes. The cancel: node is a leaf with when on a verdict other than
pass. The node that opens the pull request has when on a verdict equal to pass and carries no
trigger_rule. build-pr-gate still depends on the node that opens the pull request, so a green
build in hitl still pauses there. A reader can confirm that no node in /build carries
trigger_rule. The green path is the one most easily broken: a green build must open its pull
request.
Holds (at 579b4ef4, open-pr stages, commits and opens the pull request in one node, and no
step pushes by name). Amended 2026-10-28 by the maintainer: committing and pushing are two nodes,
and both run on every verdict. A prompt node cannot fail by reporting, so one prompt node cannot
make a failed push fail the run. A bash node cannot tell the build's source files from scratch
files. So the work splits in two, and the implementation commits and the Session artefacts are
on the remote before the verdict decides anything:
- A prompt node commits. It depends on report, because the list it stages includes
  report.md, and criterion 7 promises that file inside the pushed branch. It stages the same
  named-path list open-pr stages at 579b4ef4, now with evidence.json always on it, and
  commits. The staging rules move with the staging, inline in this node's prompt: named paths, one
  git add per path, never a sweep, never $ARTIFACTS_DIR. It does not push.
- A bash node pushes. It depends on the commit node. Before the push it exits non-zero when
  HEAD is detached or on $BASE_BRANCH, and when any of the five Session artefacts (PRD.md,
  issues.json, report.md, build-state.json, evidence.json) is missing at HEAD or differs
  from the working tree. That check is how a commit the prompt node did not make fails the run.
  It outputs the branch it pushed and the pushed commit as fields. Criterion 7's cancel: message
  references them.
- When the push or a check fails, the bash node fails, both siblings of criterion 5 are skipped,
  and the run ends failed, not cancelled. That is intended: no work reached the remote, which
  is a different fact from a red verdict.
- The node that opens the pull request keeps only opening it against $BASE_BRANCH
  (#505) and linking it to the slices'
  tickets.
What this criterion does not decide. The bash node checks the five Session artefacts only. If
the prompt node leaves a source or test file out of the commit, the push succeeds and the branch
lacks it. One prompt node at 579b4ef4 had the same gap, so the split does not widen it. Whether
to close it is an open question for #667.
New, replacing the old criterion 7, which moved. The cancel: message tells the operator
where the work is. It names the verdict, the branch the commit-and-push node pushed (from that
node's output, criterion 6), and the path of report.md inside that branch. It writes nothing to the tracker.
Moved to feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket.
to 13. Moved to feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket. Old criterion 13 (the baton
becomes an object) existed only to carry the grouping ticket's id, so this ticket leaves the baton
as it is.
Amended (claims re-found by what they say; several moved in No Box can target a base branch that is not the trunk, so run 4 cannot merge to integration/dlc #505). Every sentence that
describes the old mechanism is rewritten. Find each one by its claim, not by a line:
- verification's and goals-check's prompts say evidence withholds the file unless both
  verdicts are green, and that evidence_policy then gates on it.
- evidence's header comment says it writes the file only on green.
- open-pr's staging list says to stage evidence.json only if it exists, because evidence
  withholds it on red.
- open-pr's header comment says it always runs, even in afk, and build-pr-gate's says the
  pull request is already open above.
- The Box's description says the Box gates a pull request after report.
- docs/boxes/unic-dlc-build.md says the run still continues to report and open-pr on a
  withheld verdict, and its step for open-pr describes staging and committing there.
- The plugin AGENTS.md § Plugin doctrines says build-state.json is committed once, at
  open-pr. The rule and its reason stay. The sentence stops naming the node, so a rename cannot
  falsify it again. The Box operator doc names the new node, because it describes nodes to someone
  running them.
Holds. The evidence node's comment block keeps its warning about dollar-prefixed tokens.
The DAG loader reads every node-output token in that script body as a real edge, comments
included. This criterion is satisfied by changing nothing, and is here because rewriting the block
for criterion 2 is where the warning would be lost.
Moved to feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket.
Amended (grouping part moved). An ADR records the decision. It carries the always-write
rule, the verdict as the branch signal, the sibling shape and why the chain was not used,
evidence_policy's demotion to a backstop, and the four Archon behaviours from criterion 0 with
the probe's date and the Archon version. ADR 0034-evidence-gate-deterministic-writer § Decision
item 4 and ADR 0023-build-generic-red-green-refactor-loop's sentence that the pull request is
still opened in afk are revised inline, by the convention in the plugin's docs/adr/README.md.
Take the next free ADR number when you write it, not one named here.
Moved to feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket.
Moved to #460. The CONTEXT.md
entry for Baton sat here because old criterion 13 changed what the baton is. Without 13 this
ticket does not touch the baton, and chore(unic-archon-dlc): ticket is the plugin's word — sweep work item and issue, rename issues.json to tickets.json, reopen the Issues JSON entry #460 already carries the vocabulary sweep and the
issues.json rename.
New. The diagrams that show the changed nodes are regenerated.
docs/architecture/20260925-unic-dlc-build-pipeline.json gains the commit-and-push node and the
two siblings, and both its PNGs are exported again. The architecture and Box-set diagrams each
name one of these nodes: read them, and regenerate any whose content the change falsifies.
New. CHANGELOG.md records the change under ### Breaking. A red /build run used to
end failed with a pull request open. It now ends cancelled with none. An operator or a script
that waited for a red pull request sees a different outcome.

Two decisions that are not criteria

Both are satisfied by changing nothing. Writing them as criteria would buy coverage and cost an audit
round to discover.

The order of open-pr and build-pr-gate does not change. The title of this ticket says it
should, and the measurement says otherwise. Once a red verdict opens no pull request, the gate no
longer authorises an effect that should not have happened. It reviews work that met its criteria.
evidence_policy: { required: true } stays. With always-write it can no longer refuse a red
run. It now covers only the case the cancel: node cannot see, which is evidence itself failing
before it writes. Say that in the ADR, or the next reader finds a config key that never fires and
deletes it.

What is still decided by a config value, after this

gates.build still decides whether a human sees a green pull request before the run ends. That is
intended. On green the pull request is legitimate, so skipping the pause is a tenant's choice rather
than a hole.

What this did not decide

Nothing is left open in this ticket. The two unknowns of 2026-09-04 both concern the grouping ticket
and moved with it: whether /qa's close of the grouping ticket deserves its own gate, and what a
Consumer whose tracker expresses no relationship should do.

#652: lands first. /qa ends green
without merging when no open pull request exists.
#505: landed. open-pr targets
$BASE_BRANCH. Keep that target when the node is split.
feat(unic-archon-dlc): a red /build run reaches a human through a grouping ticket: the grouping-ticket half, split out on 2026-10-26.
Blocked by this ticket.
#460: takes the Baton entry.
#459,
#437,
#458,
#440: siblings. Absorb nothing.

ADRs this revises

Plugin-local, in apps/claude-code/unic-archon-dlc/docs/adr/, revised inline by slug:

0034-evidence-gate-deterministic-writer: § Decision item 4 in full.
0023-build-generic-red-green-refactor-loop: the sentence saying the pull request is still opened
in afk.

STALE since 2026-09-17 23:05 CEST. Superseded by the re-grilling opener posted below by unic-sdlc-orchestrator-1. This one names map #456, seat wayfinder-orchestrator-6 and #430 as pending; all three have moved. Kept for the record of the 2026-09-04 grilling.

How to start this session

Written 2026-09-03 19:10 by wayfinder-orchestrator-6. This is the grilling opener. It produces criteria on this
ticket, not code. The implementation comes later, in one fresh session that implements
#430 first and this ticket second; that session gets its
own opener once the grilling closes.

What it produces, and which skill

An amended body on this ticket: acceptance criteria that pass the four reads, a What this did not decide list,
and the ADRs the fix will revise, named by slug. Nothing in .archon/workflows/ changes in this session. A Box
YAML edit here is a defect, not initiative.

Enter with /wayfinder 456 439, typed by the maintainer. It claims the ticket by assigning it, which nothing
else does. Check the assignee first. If it is already set, report "already claimed" and go on. Then leave the
skill at its step 4 and do not close this issue. This stream works a bug under its own number; the decisions
go into this body, the maintainer closes after the implementation merges, and the orchestrator writes #456's
Decisions-so-far line then. Said again under § On landing.

The grilling is /grill-with-docs, typed by the maintainer. Its frontmatter carries
disable-model-invocation: true, so you cannot invoke it, and /grilling alone is the wrong tool: #441 ran it for
six rounds and settled decisions had nowhere durable to land. /grill-with-docs composes /domain-modeling,
which will want to write an ADR. Write no ADR file in this session. Put the decision text in this ticket
under a heading the implementer can lift, and name the ADR by slug. #430's implementation writes an ADR in the
same directory, and two sessions taking 0037 is how numbers collide. /writing-for-agents for every criterion
you draft: an implementer agent reads them.

Measure the premise first, both halves

The ticket alleges two things. Both are true as measured, and neither is what the ticket implies.

Half one, the ordering. open-pr (unic-dlc-build.yaml:590) depends on report and runs before
build-pr-gate (:671, when: gate == 'hitl', on_reject with max_attempts: 3 at :699). Verified on
0b74117. But it is not an oversight. Three documents record it as a chosen cost:

docs/adr/0034-evidence-gate-deterministic-writer.md § 4: the evidence node exits 0 on a red verdict so the
run reaches report and open-pr, "so a human gets the report and a PR-shaped surface showing what failed".
Then, verbatim: "The cost of that choice is that a red build still opens a PR — in gates.build == afk with no
build-pr-gate pause at all — so open-pr must treat a missing evidence.json as expected and skip staging it
rather than failing on git add. If that cost is judged too high, gate open-pr on the verdict; that is a
separate decision with its own blast radius, not a byproduct of this one."
docs/adr/0023-build-generic-red-green-refactor-loop.md § 6: the gate is skipped in afk "(the PR is still
opened by the preceding open-pr node)".
docs/boxes/unic-dlc-build.md step 6: "The run still continues to report and open-pr on a withheld
verdict, so you get the report and the PR. It is the run status the engine refuses."

So this ticket is the separate decision ADR-0034 deferred. The grilling reopens a recorded trade, and the
first question is whether the reason for the trade still holds: on 2026-08-10 the report existed only in the
worktree, so failing before open-pr destroyed it. Ask what else would carry it now.

Half two, the config gate. The Consumer's .archon/unic-dlc.config.yaml on origin/develop sets
gates.build: afk and gates.pr-review: afk (:87–91), with the comment the ticket quotes at :85–86. So in
run 2 build-pr-gate was skipped because the config said afk, not because of the ordering. The two halves
compound rather than overlap: in afk there is no pause at all; in hitl the pause comes after the side effect.
Each half on its own would be a smaller ticket. Say in the criteria which half each one addresses.

Read order

This ticket's body, then register rows 39 and 40 in
docs/research/2026-08-31-dlc-run-2-findings.md (line 139 onward). Row 39's evidence column says
"verified in the DAG"; the three documents above are the part it does not say.
ADR-0034 § 4 in full, then ADR-0023 § 6, then the build Box doc step 6 and step 8. These are the decision
you are reopening. Quote them in the body, do not paraphrase.
#456 § Notes, the class rule. Read it before you
draft one criterion; see § The decision the ticket does not make, item 1.
The engine's approval schema: docs/adr/0011-archon-schema-target.md:43. An approval: node carries
message, optional capture_response, and on_reject { prompt, max_attempts }. There is no on_approve.
No Box uses capture_response (grep -rn capture_response apps/claude-code/unic-archon-dlc/.archon/workflows/
returns nothing). A reject with no on_reject cancels the run; a reject past max_attempts cancels the run.

The four gates, as they stand on 0b74117, in apps/claude-code/unic-archon-dlc/.archon/workflows/:

Box	Order	Reject behaviour
`/build`	`open-pr` (`:590`) → `build-pr-gate` (`:671`)	`on_reject` re-verifies and re-pauses, `max_attempts: 3`. The PR already exists.
`/pr-review`	`review-gate` (`:536`) → `post` (`:560`, `trigger_rule: all_done`)	no `on_reject`: a reject cancels and every finding is lost. Gate before side effect, and it is the one the maintainer could not use.
`/qa`	`uat-gate` (`:211`) → `verify-pr-base` → `merge-gate` (`:320`) → `merge` (`:334`, `when` fail-closes at `:337`)	`uat-gate`'s `on_reject` (`:220`, `max_attempts: 1` at `:267`) files every finding on the tracker before the second reject halts. The one gate whose reject preserves work.
`/explore`	`spike-ticket` (`:356`) → `spike-branch-gate` (`:415`) → `preserve-spike`	Ticket filed before the gate "so the durable output survives a discard" (comment at `:350–355`, ADR-0029 § 4). No `on_reject`; a reject cancels with the ticket already filed.

Two Boxes already answer this ticket's second question for one artefact each. Neither answers it for a
pull request.

Archon facts that bound every answer, in the Consumer's docs/agents/unic-archon-dlc.md on
origin/develop (git show origin/develop:docs/agents/unic-archon-dlc.md, read the file there and never a
main checkout): § Three Archon defects that land here, and § Recovering a killed run. Terminal approve
approves and resumes; a UI approve parks the run forever; resume walks through an unresolved gate and
posts; abandon is the kill; archon complete <branch> --force deletes worktree and branch, so a
/build run abandoned before open-pr loses its commits unless someone pushed them. Tool inputs survive
in ~/.archon/archon.db, so a killed run is recoverable by hand. Whether a design may lean on that is a
grilling question, not a given.
The gate vocabulary, two surfaces: gates.<box>: hitl | afk for the four Boxes (commands/setup.md:271,
README.md:183, ADR-0017 § Consequences), and specs.gate / tickets.gate: open-pr | stage-only for the two
commands (README.md:179–180). The ticket's "no safe middle" is about the first. Ask whether the second
already has one, and whether stage-only is the shape.
apps/claude-code/unic-archon-dlc/AGENTS.md § Plugin doctrines, the /qa and /pr-review bullets (lines
92–93), and CONTEXT.md for Evidence gate, Evidence set, Holds. If the grilling coins a term for
the thing that must survive a stop, it goes into CONTEXT.md in the implementation, and the body names it now.
Root AGENTS.md § Acceptance criteria are prose. The four reads, run on your criteria before you post them.

The one trap that costs an hour

"Reverse the order" is not one change, and it recreates the loss it is meant to prevent. Put build-pr-gate
before open-pr and the run's whole product, the commits on the branch, report.md, evidence.json, exists only
in an Archon worktree until a human approves. A reject cancels the run. archon complete <branch> --force, the
sanctioned cleanup, then deletes the worktree and the branch. That is the #438 shape of loss, on a leg that
took six points of a week to produce, and ADR-0034 chose the current order to avoid exactly it. So the grilling
has to separate the durable output (commits pushed to a branch; report.md) from the outward-facing side
effect (a pull request a client can see and a /qa run can merge), and decide the order for each. /explore
did that split for a ticket and a branch. Nobody has done it for a PR.

The decision the ticket does not make

Five places where two competent people diverge. The grilling decides each, or lists it as undecided.

Which class this fix is, under wayfinder: run 3 — do mechanisms carry the knowledge where documents did not? #456's rule. An instrument changes whether a result can be read; a second
variable changes what the chain produces. A fix that stops a red build from opening a PR changes what run 3's
build leg produces, and its effect on the run is not separable from the mechanism checks under test. If the
grilling lands there, this ticket leaves run 3's map: Dispatch run 3, read both legs against sealed predictions, and write it up #457 loses a blocker, the fix joins run 4's set, and
ADO 43036 moves with it. A fix that only changes what
survives a stop, and what the human at the gate can read, stays an instrument. Rule this first; every
criterion depends on it.
Who stops the run, and what must survive each stop. Three stoppers, three answers. A human at a hitl
gate who wants to read before deciding: every finding and every commit must survive, and the gate must not
be a choice between "post it" and "destroy it". An operator killing an afk run: the side effect may already
exist, so the answer is what the run leaves behind, not what it withholds. The engine failing the run after
the side effect, run 2's shape: whether evidence_policy may refuse completed on a run whose PR is already
open, or whether a withheld verdict gates open-pr itself, which ADR-0034 named and did not take.
Whether a pushed branch without a pull request is a side effect. It is visible on the host and it is not
reviewable or mergeable. If it counts as durable output and not as a side effect, the /explore split
applies to /build directly: push before the gate, open the PR after. If it counts as a side effect, nothing
durable can precede the gate and item 6 in the read order decides how much loss the design accepts.
What "a safe middle" is, given the schema. No on_approve. A gate that pauses and lets a human read
without a destroy-on-reject needs an on_reject whose prompt does nothing but re-pause, bounded by
max_attempts, or a capture_response nobody has used. Both are workarounds dressed as design. A third
gates.<box> value is the other route, and the ticket's own claim is that a config value cannot be the
middle. Test that claim against specs.gate: stage-only, which is a config value and is a middle.
Which Boxes the fix reaches. The ticket names /build and the pr-review config comment. review-gate
has no on_reject and destroys findings on reject; uat-gate already preserves them. Say whether the
criteria cover one Box, the two with a pull-request side effect, or all four, and say why the ones left out
are left out. A criterion that says "every Box" and names one is the vacuous shape.

Other traps

bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430 lands first, and it edits the same files. bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430's criteria 8 and 9 change /qa's merge when:
(unic-dlc-qa.yaml:337) and add an unresolved state to evidence (unic-dlc-build.yaml:472–560). The
maintainer's plan is one session implementing bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430 then bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439. Write bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439's criteria against the tree bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430
leaves, not 0b74117, and say so in the body: "after bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430, merge's when: reads …". A criterion written
against today's line numbers is stale before it is dispatched.
feat(unic-archon-dlc): the /specs gate asks for approval of an artefact nobody reads #437 is a sibling, not a part. The /specs gate approves an artefact nobody reads; it is a command gate,
not an Archon approval: node, and it was blocked by fix(unic-archon-dlc): /specs skips the grilling its config requires and reports halts that never fired #441, which merged on 2026-09-03. Name it in § Related
and leave its criteria alone.
bug(unic-archon-dlc): three /qa nodes emit prose that nothing reads and nothing keeps #458 owns /qa's orphan outputs and is outside run 3. If the grilling finds that "what survives a stop"
in /qa depends on e2e.summary or uat-prep.criteria reaching a file, that is bug(unic-archon-dlc): three /qa nodes emit prose that nothing reads and nothing keeps #458's finding. Point at it,
do not absorb it.
bug(unic-archon-dlc): the re-entry idiom reads before it offers the choice, and Step 8 hardcodes names the Consumer overrides #440 owns the re-entry idiom and Step 8's names. Out of run 3. Same rule.
The plugin names no tool and no host. A criterion that says "open the PR through gh" or "push with
git push" fails the doctrine at AGENTS.md:61. Name a role and a file; the node reads the tracker contract.
ADRs are revised inline, by slug. The dated > **Amended** blockquote is retired since docs(unic-archon-dlc): ADR-0020 states superseded claims below three amendment blocks #452. The ADRs
this fix revises: 0034-evidence-gate-deterministic-writer § 4, 0023-build-generic-red-green-refactor-loop
§ 6, 0025-qa-pipeline-onramp § 3–4, 0026-pr-review-generic-archon-harvest § 6, 0029-explore-research-spike-onramp
§ 4, and the gates line in 0017-container-follows-structural-need § Consequences if the value set changes.
The body lists them; the implementation edits them.
Absolute paths in every Bash call; cd does not persist between calls here. Never a foreground sleep.
You edit no repository file in this session, so no worktree is needed; if /domain-modeling insists on
writing one, stop and put the text in the ticket instead.
Both develops are current at 0b74117 (plugin) and e74778e (Consumer) as of 19:00. Re-run
git fetch and git log -1 origin/develop in each before quoting a line number.

The bar that proves it

The false pass: a body with criteria that reorder two nodes in unic-dlc-build.yaml, plus one that says
"a gate value has a safe middle", plus nothing about what happens on reject, abandon or engine failure. That
set reads as coverage. It would ship a /build whose reject destroys six points of work, and it answers neither
of the ticket's two questions.

The real bar: for each of the three stoppers in item 2, the body says what survives and where it lives after
the stop, in one sentence each, and at least one criterion is able to act: it names a thing a node must do
(file, push, write, re-pause) rather than a thing that must not happen. Then the four reads, as one set: is there
a gap between two criteria an implementer fills with a defensible choice; what is still decided by a config value
after the fix; which criterion passes on the current tree unchanged; and what did the grilling not decide. Write
the last list into the body under What this did not decide, numbered, so the implementation opener can point
at it.

On landing

Amend this body: criteria, § Related (feat(unic-archon-dlc): the /specs gate asks for approval of an artefact nobody reads #437, bug(unic-archon-dlc): three /qa nodes emit prose that nothing reads and nothing keeps #458, bug(unic-archon-dlc): the re-entry idiom reads before it offers the choice, and Step 8 hardcodes names the Consumer overrides #440, bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430 with "lands first"), the ADR slugs, What
this did not decide, and the class ruling from item 1 in one sentence at the top. Say on the ticket that the
body was amended and when.
Do not close this issue. Keep the number; ADO 43036 points at it. The maintainer moves needs-specs to
ready-for-agent when he approves the criteria, or asks for another round. The orchestrator writes wayfinder: run 3 — do mechanisms carry the knowledge where documents did not? #456's
Decisions-so-far line only when the implementation merges.
If item 1 rules this a second variable, say so to the orchestrator before anything else; the map changes, and
that is the orchestrator's edit, not yours.
No version bump, no changelog, no PR. This session ships prose on a ticket.

Report back

BEFORE YOU FINISH, AND WHENEVER YOU LEARN SOMETHING THAT WOULD HAVE CHANGED THIS PROMPT, SendMessage to
the live orchestrator seat. ListAgents names it — the seats are wayfinder-orchestrator-<n> and the
number rises with each restart; it was wayfinder-orchestrator-7 when this line was last checked, on
2026-09-05. Read the name from ListAgents, never from this line. If no orchestrator seat is live, comment
these four things here instead:

what you expected
what happened
the command or file that proved it
whether it changes only this ticket, or every session in this repository

Send it when you learn it, not only at the end. A session that dies unreported takes its learnings with it.

Body amended 2026-09-04 21:30 CEST, after a grilling with the maintainer under /grill-with-docs. The two original questions are replaced by seventeen acceptance criteria, two decisions recorded as decisions rather than criteria, and a What this did not decide list of five items.

Three things a reader of the old body should know before reading the new one.

Both original claims failed measurement. The node ordering did not cause run 2's outcome — gates.build was afk, so the gate was skipped by its own condition wherever it sat. And a reject at /pr-review's review-gate does not destroy the findings: four files are on disk before the gate pauses. What is left is the real defect, that a build failing its own acceptance criteria opens a pull request, which is ADR-0034 § Decision item 4's deferred trade.

Three Archon behaviours were measured rather than assumed, in a throwaway repository outside every clone, six runs, worktrees removed afterwards. A cancel: node sets terminal status cancelled, distinct from failed, and stops the DAG between layers. A node after a conditional cancel: is skipped on the non-cancelled branch unless it carries trigger_rule: all_done — so without that rule a green build would open no pull request. And a script: node's output_format is ignored, its stdout returned raw, so evidence must print JSON for anything to branch on its verdict. Criteria 2 to 5 depend on all three.

Class ruling: second variable. This changes what /build produces and what /tickets creates, and neither effect separates from the mechanism checks run 3 measures. #439 has left #457's blockers.

Two spin-outs were approved and are not filed yet: one for /pr-review's missing on_reject plus the cleanup.prune_slug_dirs default, one for the vocabulary ruling in What this did not decide item 3. The label stays needs-specs until the maintainer approves the criteria.

Four reads on the seventeen criteria, 2026-09-04 21:40, wayfinder-orchestrator-6. Two findings, both applied to the body inline; the rest holds.

Gap between criteria 4 and 7. The cancel: node stops the DAG between layers; criterion 7 named no node for the pointer and no position. A pointer node beside or after cancel: never runs and the red path passes 4 to 6 with 7 failing silently. Criterion 7 now names the node and its edges: after commit-and-push, before cancel:.
A fourth surface criterion 6 falsifies. "build-state.json is committed once, at open-pr" lives in the plugin AGENTS.md staging doctrine and in the build Box's operator doc. Criterion 14 now names the claim.
Verified, no change: /qa, /pr-review and /explore do not read issues.json, and commands/triage.md names the file twice without asserting its shape, so criterion 13's six files are the whole set. Criteria 8 and 15 pass on an unchanged tree and the body already says so.

Spin-outs filed: #459 (review-gate re-pause, unresolved gate path, prune_slug_dirs default) and #460 (the vocabulary question). Label stays needs-specs until the maintainer approves the set.

Second grilling pass, 2026-09-04 22:45. Four of the five What this did not decide items were ruled and the criteria now run to nineteen.

18 — /qa closes the grouping ticket when it merges, and gains a baton read in bootstrap. Measured while writing it: /qa, /pr-review and /explore read no baton today; only /build does.
19 — CONTEXT.md gains an entry for Baton. The word appears 32 times across the plugin and CONTEXT.md defines it nowhere, which makes the most-used structural term also the most-used undefined one. Its _Avoid_ line names handoff contract, because AGENTS.md already calls one baton exactly that — and says in the same breath that /handoff and HANDOFF.md keep their names, so a later sweep does not rename one of Matt's skills.
12, extended — /tickets accepts the grouping ticket as an optional argument. Work often already lives under a ticket a human made; this repository's own effort is four stories under one Azure DevOps feature, created by hand before any Box ran.
14, sharpened — AGENTS.md keeps the build-state.json rule and its reason and stops naming the node. The rule survives a rename; at open-pr does not, which is the only reason criterion 6 falsified it. The Box operator doc names the new node, because it describes nodes to someone running them.

The word is ruled: ticket. work item is Azure DevOps's, issue is GitHub's and Jira's, and ticket is the one no host in front of us claims. #460 carries the sweep, the argument CONTEXT.md needs against its current _Avoid_: tickets line, and the issues.json → tickets.json rename.

Two items remain undecided, both honest unknowns rather than unasked questions: whether /qa's close deserves its own gate, and what a Consumer whose tracker expresses no relationship at all should do — criterion 11 stops there, which is correct and untested, no such tracker having been met.

docs(unic-archon-dlc): record that the build verdict decides the pull…

docs(unic-archon-dlc): name the evidence_policy backstop as ADR-0040 …

docs(unic-archon-dlc): draw the verdict siblings in the build pipelin…

fix(unic-archon-dlc): refuse to push uncommitted or base-branch work …

docs(unic-archon-dlc): export the build pipeline PNGs again

fix(unic-archon-dlc): check every Session file and name the resume ca…

docs(unic-archon-dlc): say push-branch refuses a resumed run only whe…

Criterion 6 amended, 2026-10-28 (maintainer decision, recorded by unic-sdlc-orchestrator-8)

Criterion 6 asked for one node that commits and pushes, and its 2026-10-26 amendment asked for a
failed push to end the run failed. One prompt node cannot do both: a prompt node that reports a
failure still finishes green. PR #665 splits
the work into commit-session (prompt) and push-branch (bash), and the maintainer accepts that
shape. The criterion now describes two nodes, so the ticket and the merged code agree.

The criterion now also names the one gap the split leaves: push-branch does not check source or
test files. That gap existed before the change. See the criterion's last paragraph and #667.

Nothing else in the body changed.

Navigation Menu

Repository navigation

bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439

What this is

The decision

Rulings of 2026-10-26

Acceptance criteria

Two decisions that are not criteria

What is still decided by a config value, after this

What this did not decide

Related

ADRs this revises

How to start this session

What it produces, and which skill

Measure the premise first, both halves

Read order

The one trap that costs an hour

The decision the ticket does not make

Other traps

The bar that proves it

On landing

Report back

22 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439

Description

What this is

The decision

Rulings of 2026-10-26

Acceptance criteria

Two decisions that are not criteria

What is still decided by a config value, after this

What this did not decide

Related

ADRs this revises

Activity

orioltf commented on Sep 3, 2026

How to start this session

What it produces, and which skill

Measure the premise first, both halves

Read order

The one trap that costs an hour

The decision the ticket does not make

Other traps

The bar that proves it

On landing

Report back

orioltf commented on Sep 4, 2026

orioltf commented on Sep 4, 2026

orioltf commented on Sep 4, 2026

22 remaining items

orioltf commented on Oct 28, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions