Terminal lane: repairs after the cross-family review (gpt-6.1-sol through the packaged Codex lane on OmniRoute), 46 findings checked, 39 repaired - #532

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…ine sweep, tmux shutdown by identity, complete scan, landscape and colour readers, scoped docs)

Six read-only jobs of the packaged Codex lane (three lenses, each on gpt-6.1-sol and on gpt-6-astra-ultra, effort max, through the local OmniRoute gateway) read the repairs of #532 and returned 37 findings. 21 were graded by Claude verifiers (17 confirmed, 4 partly, none refuted), 16 are duplicates, and all 37 are repaired here.

- pty probes: the signal handler only latches the first signal; cleanups run to their end at safe points and the exit status 128+signal follows; the selftests deliver a SIGTERM at every line of the resource lifecycle; pty_probe_mutants.py has seven mutants per probe (8 runs each, 0 problems)
- tmux probes: the private server is verified gone by pid, start time and socket command line before its directory is removed; two shutdown-failure controls each
- the notification-type scan, the landscape comparison and the colour counter follow their sources' grammar (complete initializers, every declared catalog, every Default and Display impl and the Tui fields, the pinned Windows Terminal SGR rules)
- documentation says what each check proves (the nonce does not authenticate stdout, the digest prefix is recorded, the real-/tmp count varies); the record builders read private labels from a private file
- sanitized record of the round, receipt rebuilt with the final read, correction notes in three older receipts, four anti-pattern rows, decision record subsection

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…ine sweep, tmux shutdown by identity, complete scan, landscape and colour readers, scoped docs) (#552)

Six read-only jobs of the packaged Codex lane (three lenses, each on gpt-6.1-sol and on gpt-6-astra-ultra, effort max, through the local OmniRoute gateway) read the repairs of #532 and returned 37 findings. 21 were graded by Claude verifiers (17 confirmed, 4 partly, none refuted), 16 are duplicates, and all 37 are repaired here.

- pty probes: the signal handler only latches the first signal; cleanups run to their end at safe points and the exit status 128+signal follows; the selftests deliver a SIGTERM at every line of the resource lifecycle; pty_probe_mutants.py has seven mutants per probe (8 runs each, 0 problems)
- tmux probes: the private server is verified gone by pid, start time and socket command line before its directory is removed; two shutdown-failure controls each
- the notification-type scan, the landscape comparison and the colour counter follow their sources' grammar (complete initializers, every declared catalog, every Default and Display impl and the Tui fields, the pinned Windows Terminal SGR rules)
- documentation says what each check proves (the nonce does not authenticate stdout, the digest prefix is recorded, the real-/tmp count varies); the record builders read private labels from a private file
- sanitized record of the round, receipt rebuilt with the final read, correction notes in three older receipts, four anti-pattern rows, decision record subsection

Co-authored-by: Scout <scout@local>
Co-authored-by: Claude Sonnet 5.5 <[email protected]>

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…portable Claude template and the Codex AGENTS template, anti-pattern rows and the folded finalization record (unit F1) (#557)

* Codex user template: standing rule clauses, skills, models, routing and token lanes

Adds the 2026-09-30 standing clauses to the top-rule block of
adoption/templates/codex.AGENTS.template.md: skill discovery (search-first,
find-skills, skill-creator; implicit invocation from a skill's description and
explicit $skill-name, per openai/codex rust-v0.159.2
codex-rs/ext/skills/src/catalog_prompt.rs:8), upstream A/B and E2E harnesses,
the completeness critic feeding the next landscape sweep, the north-star
direction, GPT-6 model routing through the OmniRoute gateway (codex -p
omniroute), dated decision records, the startup rule, and a token-lanes line
naming each MCP server the Codex config template registers.

The RTK upstream text and the exceptions block stay byte-identical, so the F4
block in every Codex role is unchanged. The top-rule pin is re-derived with
template_segments(): 455 words, 7b41478f...; a new test requires a lane for
every server in codex.config.template.toml and the standing phrases.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Portable Claude template: superset of the user-level rules plus the standing clauses

examples/claude-native/CLAUDE.md becomes the single managed source of the
operator's user-level file. Step 1 of the top rule takes the user-level
paragraph (research and record, reference implementations, popularity guides
discovery) and the skill-discovery clause (search-first, find-skills with
`npx skills find`, skill-creator); the core rule takes the rules only the
user-level file held (short plan, pins and reasons, relevant layers, reviewer
agreement is not proof), the north-star direction, the upstream A/B and E2E
harnesses and the completeness critic; token practice takes the startup rule;
the worker section takes GPT-6 routing through the OmniRoute gateway. Every
worker, model, Ultracode and agent-team rule is kept word for word.

PortableTopRuleTests gains a phrase check for the clauses and the user-level
rules (it failed on the unedited template with 29 phrases missing, 8 of them
rules the user-level file held) and re-baselines the word budget to 1,656.
@RTK.md stays the host's own import (recipes/claude-native-profile.md:134).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Decision record: standing rule text in every instruction layer; divergence log row

docs/decisions/2026-09-30-rule-text-every-layer.md supersedes the 2026-09-28
top-rule record's "the operator's user-level file is the operator's own" at
the user's 2026-09-30 request: the portable template becomes the single
managed source of that file. It maps the six clauses to each surface, records
the alternatives (a UserPromptSubmit carrier, per the hooks reference, adds
its context beside every prompt; leaving the file unmanaged produced the
divergence; per-agent copies are unit F2's; pointers measured as the
fallback), the o200k sizes from token_manifest.count_files, the overrun of
the unit's 200-token budget with its floor and fallback, and the overturn
conditions.

docs/harness-defaults.md logs the host/template divergence, naming only the
check seen failing on it (the new PortableTopRuleTests phrase check).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* AGENTS.md: the six standing clauses on the four named lines; re-register hashes

Hot files last (docs/lanes.md hot-file protocol). AGENTS.md takes the
standing clauses by amending, not appending, the lines the brief names:
- line 3, the top rule: skill discovery (search-first, find-skills with
  `npx skills find`, skill-creator; manifest skills model-invocable in both
  clients) replaces "with the installed research and skill-discovery
  skills", and A/B and E2E on upstream harnesses, never a self-written runner;
- line 7: the north-star R&D direction; each unit names its action;
- line 18: the completeness critic feeding the layer's next landscape sweep;
- line 28: GPT-6 through the OmniRoute gateway (astra max, sol medium, 6.1
  sol pending qualification), Codex CLI as the second native client, and "no
  audits, trials or network at startup" with the daily currency timer's
  due-file line (record lands with unit A2) replacing "Do not rerun the full
  audit or model trials at startup".

manifests/evidence.json re-registers the six changed files already listed,
with scripts/host_receipts.py register_file; component_matrix and
new_host_grand_list --check pass unchanged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record and log row: nine user-level rules, not eight

The phrase check's committed form (with "from the selected source revision",
the user-level file's own wording) finds 29 phrases missing from the unedited
template, 9 of them held only by the user-level file; the record, its Checks
section and the anti-pattern row said 8 (from a run before that phrase
changed). The record also states the two non-verbatim worker lines exactly:
"Quality comes first" is one line with an identical word sequence, and the
sizing line keeps every user-level word and adds "to its task".
Re-registers docs/harness-defaults.md (hot file, last commit).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Codex user template: Sol-primary routing and skill matching folded into the standing lines

Folds the two lines the main checkout's uncommitted changes added inside the
top-rule block (pre-existing uncommitted changes observed in the main
checkout; original author not established), merged with this unit's lines
instead of carried beside them:
- the Models line now states the Sol-primary routing of
  docs/decisions/2026-09-30-sol-primary-quality-defaults.md (gpt-6.1-sol at
  ultra for coordination and at max for workers, gpt-6-astra at max after the
  recorded escalation triggers), replacing the superseded astra/sol-medium
  wording and "gpt-6.1-sol waits for qualification"; the OmniRoute gateway
  clause stays;
- the skills line takes the skill-matching rule (read each selected SKILL.md,
  follow its native workflow, load references only when needed) next to
  implicit and $skill-name invocation.

The phrase test failed first on the unedited block with five phrases missing
(exit 1). Pin re-derived with template_segments(): 496 words, 2d3107a2...;
the RTK text and exceptions block are unchanged (test_codex_agents 18 OK).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Portable Claude template: Sol-primary Codex routing; skill matching kept inside "Keep context small"

- The GPT-6 bullet follows docs/decisions/2026-09-30-sol-primary-quality-defaults.md:
  Codex CLI is the second native client, gpt-6.1-sol at ultra coordinates and
  at max runs workers, gpt-6-astra at max takes the recorded escalation
  triggers; cross-family research, review and sweep votes run through the
  OmniRoute gateway. It replaces the superseded astra/sol-medium wording and
  "gpt-6.1-sol pending qualification".
- Folds the skill-matching wording of the main checkout's uncommitted change
  (pre-existing uncommitted changes observed in the main checkout; original
  author not established) into the token-practice bullet. The change as
  observed replaced "Keep context small", a rule the operator's user-level
  file holds; the merged bullet keeps both.

The phrase check failed first on the unedited template with four phrases
missing (exit 1); it now also requires "Keep context small", which the
observed version lacks (mutant check: 31 phrases missing there, that one
included). Word baseline 1,656 -> 1,680.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* AGENTS.md: fold the skill-lifecycle and Codex-defaults bullets; one model-routing sentence set

Folds the two bullets of the main checkout's uncommitted AGENTS.md change
(pre-existing uncommitted changes observed in the main checkout; original
author not established). Neither was on origin/main at 11227bf (the file's
blob there equals e45328d's) nor on any other unit branch:
- the token-practice bullet on matching skill descriptions, reading each
  selected SKILL.md and adoption/skills/lifecycle.md (which unit F3 adds);
- the Codex defaults bullet (gpt-6.1-sol / ultra coordinator, gpt-6.1-sol /
  max primary workers, Astra/max on the recorded triggers, the Sol-primary
  record that unit D4 adds), merged with this unit's clause: Codex CLI is the
  second native client, and cross-family research, review and sweep votes run
  through the OmniRoute gateway. The startup bullet keeps only the startup
  clause, so the model routing is stated once.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Convergence guide and token handbook: Ultracode decoupled from effort; Sol worker command

Folds the true delta of two files from the main checkout's uncommitted
changes (pre-existing uncommitted changes observed in the main checkout;
original author not established), taken as a three-way merge onto
origin/main so main's own later edits stay:
- docs/convergence-architecture.md: Claude Code 2.1.284 decoupled Ultracode
  from effort; xhigh (saved fallback) and max (launcher) are separate choices,
  not measured quality gains. Source: the tagged anthropics/claude-code
  v2.1.285 CHANGELOG.md line 206 ("it no longer forces xhigh effort and stays
  on at any effort level", under 2.1.284), read 2026-09-30.
- docs/token-session-handbook.md: the Codex worker command names
  gpt-6.1-sol, matching the stack-worker profile and recipe that unit D4
  changes in the same batch.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Anti-pattern log: ten folded rows and six Codex runtime lane rows

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Fold the 2026-09-30 native practice finalization record, attributed

docs/decisions/2026-09-30-sota-native-finalization.md is folded unchanged
from a read-only snapshot of the main checkout (pre-existing uncommitted
changes observed in the main checkout; original author not established;
tracked diff sha256 314bd1b260da0939). One attribution paragraph after the
title states that provenance and which linked evidence is not published with
it: evidence/artifacts/sota-finalization-20260930/ has no assigned owner, the
native-skill-finalization artifacts land with unit F3 and the
codex-01592-qualification artifacts wait on unit D4. The body is
byte-identical to the snapshot (12,956 bytes); the privacy scan found no host
path, home directory or host user name in it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: rebuilt base, Sol-primary clause (a), folded items, lane rows and guards, re-measured sizes

- Base moves to origin/main@11227bfd (the three surfaces are byte-identical
  to e45328d's).
- Clause (a) follows docs/decisions/2026-09-30-sol-primary-quality-defaults.md
  (unit D4); the brief's first wording and its 04:35Z "Astra at ultra for
  complex workflow tasks" relay are recorded as superseded by that later,
  more specific user selection.
- Lists every item folded from the main checkout with the provenance
  sentence, the snapshot hashes, the three-way-merge method and what was not
  folded (recipes/README.md is on D4's branch; the finalization evidence
  directory has no owner).
- Records the six Codex runtime lane rows (relays and what was verified) and
  the three relayed guards: #543 not described as published without a native
  gh lookup that exits 0, the 07:34Z/06:46Z actor unknown, a historical OSV
  pass not current after #546.
- Sizes re-measured with tools/token-report/token_manifest.py count_files
  (gpt-tokenizer 4.0.0): AGENTS.md +323, portable template +363 (sum +686
  against the 200 budget: +139 folded text, +547 clauses and user-level
  rules), Codex template +507.
- Limitations updated: clause (c) against F3's settings, cross-unit
  references, unpublished evidence, older Codex pins, A3's scaffold copy.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Codex routing on all three surfaces: Astra/Ultra coordinates a complex workflow, Astra/Max takes a single consequential judgment

Follows the coordinator's decision and unit D4's refined Sol-primary record
(claude/sota-defaults-d4-codex-0159-20260930 at 9dd4ebe), which splits the
escalation by the Codex catalog semantics it documents (Ultra: proactive
delegation with the model's xhigh reasoning; Max: the highest reasoning
effort) and quotes the user's 2026-09-30 selection ("astra ultra when tasks
needed suitable for complex workflow"):
- gpt-6-astra at ultra when a complex workflow needs Astra to coordinate it;
- gpt-6-astra at max for a single consequential judgment (conflicting primary
  evidence, consequential architecture, complex changes across systems, or a
  failure unresolved after one bounded Sol repair).

AGENTS.md's folded Codex bullet, the portable template's Codex bullet and the
Codex block's Models line now state the split. Both phrase checks failed
first with four phrases missing each (exit 1). Codex pin re-derived with
template_segments(): 515 words, d1195686...; portable word baseline 1,697.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* One-pass tightening of the added rule text (no rule removed)

The coordinator accepted the growth over the 200-token budget and asked for
one pass that cuts words carrying no rule, limited to text this branch adds:
- AGENTS.md Codex bullet: the Astra split as two clauses and the contract
  pointer in parentheses (2,742 -> 2,733 o200k tokens);
- portable template: "(registry: `npx skills find`)" -> "(`npx skills find`)"
  (2,436 -> 2,433);
- Codex block: implicit/explicit invocation in one clause, "use the
  `search-first` skill" -> "use `search-first`", the registry label, and
  "the paired benchmark of Claude's `skill-creator` plugin" -> "Claude's
  `skill-creator` paired benchmark" (1,360 -> 1,346).

Every phrase of both phrase checks still holds. Codex pin re-derived with
template_segments(): 507 words, 97bbeb8c...; portable word baseline 1,696.
Counts: tools/token-report/token_manifest.py count_files, gpt-tokenizer 4.0.0.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Fold the finalization record's evidence: 18 of the 19 files of evidence/artifacts/sota-finalization-20260930/, unchanged

Allowed-path amendment by the coordinator. The files are copied
byte-for-byte from the read-only snapshot r2 of the main checkout
(pre-existing uncommitted changes observed in the main checkout; original
author not established); sha256sum -c against the snapshot passes.

Sanitization: nothing needed stripping and no file was changed.
- validate.py --scan-file over all 19 snapshot files: {"scanned_files": 19,
  "status": "passed"} (UUID, personal home path, Windows user path and token
  patterns).
- A wider scan found no personal home path, host user name, /tmp/claude-*
  path, session UUID, e-mail, IP address or token. The one "/home/" string
  earlier reported as a home path is the literal placeholder "/home/example"
  inside claude-review.json's prose, which validate.py's pattern exempts.
  Replacing it with <home> would alter the retained review without removing
  any host data, and would break the sha256 pin that the lane's convergence
  record holds for that file, so the file stays byte-exact. The
  config-key hits in convergence.json are recorded codex exec command lines
  already redacted to <private-path>; the three "~/." strings are generic
  install locations.

Held back: convergence.json. It is a kind "convergence_experiment" record
whose frozen inputs and observations pin five files of
native-skill-finalization-20260930 (unit F3; all five pins match F3's branch)
and three of codex-01592-qualification-20260930 (on no branch yet) by path and
sha256. validate_convergence.py --all-recorded, which CI runs, fails
discovery for an undeclared hash-listed record and fails validation for a
declared one with missing files, so it can be folded, declared in
convergence_records, only after F3 lands and D4 publishes those three files
unchanged at those paths.

The folded record's attribution paragraph says so; its links and the four
folded log rows into this directory resolve. Hash registration follows in
the last commit.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: Ultra/Max split, accepted budget miss with its overturn condition, evidence fold and sanitization, post-A3 step

- Base is origin/main@8fc86119 (rule surfaces byte-identical to 11227bf
  and e45328d).
- Clause (a) records the split from unit D4's refined record (9dd4ebe):
  gpt-6-astra at ultra when a complex workflow needs Astra to coordinate it,
  at max for a single consequential judgment, with the catalog semantics and
  the user's 2026-09-30 words as relayed.
- Measured size: AGENTS.md 2,392 -> 2,733 (+341), portable template
  2,051 -> 2,433 (+382), Codex block 829 -> 1,346 (+517); budget sum +723
  against 200, accepted by the coordinator because the clauses are the user's
  explicit directive. Composition: fold +139, split +49, tightening -12.
  Overturn condition: a rerun of the 2026-09-29 user-prefix A/B showing no
  rule-following gain for the added cost brings back the pointer fallback.
- The finalization record's evidence: 18 of 19 files folded unchanged; the
  sanitization section gives the scan and corrects the first handoff (19
  files, not 21; the one "/home/" string is the /home/example placeholder,
  kept byte-exact, also because the lane's convergence record pins its
  sha256). convergence.json is held back: a convergence_experiment record
  pinning F3 and D4 files that CI's validate_convergence --all-recorded
  would reject on this branch whether declared or not.
- Known residual links: seven targets, four from F3 and three from D4
  (the Sol-primary record and two codex-01592-qualification files).
- Post-A3 step: after #545 merges, whichever lands second copies the
  top-rule block into adoption/scaffold/AGENTS.md; not created here.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule text: conditional skill discovery on the three surfaces; the growth waiver and the provenance wording cited to the brief's amendments

find-skills (npx skills find) runs only when no listed skill fits the task, so a measured child with a fitting
listed skill makes no extra Skill or Bash call (the Gate A owner's note); search-first stays before custom code
or a tool choice. The record cites the coordinator's dated brief amendments for the accepted +723 o200k growth and
for the neutral fold provenance (a process census of file writes does not establish authorship; the Codex runtime
lane asked for the wording). Codex top-rule pin re-derived: 514 words, fecf92dc...; portable baseline 1,703 words.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

* Standing rule text: one wording on the three surfaces, coordinator-scoped (Gate A owner's review of #557)

Applies DECISION 2 of the Gate A owner's ACCEPT-WITH-CHANGES review on
AGENTS.md, examples/claude-native/CLAUDE.md and the Codex block, with the same
sentences on all three (the Codex block names a bounded worker where the
Claude surfaces name a delegated child in the skill sentence):
(a) "A coordinator, not a delegated child, invokes `search-first` before
    custom code or a tool choice; when no listed skill fits the task, it
    discovers one with `find-skills` and verifies or A/B-tests it with
    `skill-creator`."
(b) Sol/Astra routing kept for unpinned work; added: where a launch pins
    the model and effort (`-m`, `-c model_reasoning_effort`), children inherit
    that pin and a spawn call names neither; a coordinator, never a
    delegated child, starts a cross-family lane.
(c) Dropped "every manifest skill stays listed for model invocation in both
    clients" (false at this head: find-skills, grill-me and
    improve-codebase-architecture are user-invocable-only, find-skills has
    codex_enabled false) and the bare `npx skills find`.
(d) The north-star, completeness-critic, trigger/acceptance and dated
    decision-record imperatives are a coordinator's.
(f) AGENTS.md's lifecycle pointer says "(lands with unit F3)"; the
    currency-notice (A2, in the stack below) and Sol-primary (D4, on main)
    citations stay.
The harness and startup sentences take one wording too.

Post-A3 step done: adoption/scaffold/AGENTS.md carries the final top-rule
block byte for byte (tests.test_scaffold_repo passes).

Tests: StandingRuleSurfacesTests (new) asserts the shared sentences on all
three surfaces and the dropped clause on none; it failed first with 21
failing subtests (exit 1). Both phrase lists failed first (exit 1). Codex pin
re-derived with template_segments(): 538 words by Python str.split(),
9565217f...; portable baseline 1,750; comments now name str.split(), not
wc -w (review point (e)).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token handbook: the Sol worker command is marked "after D4"; the sealed E2E route keeps gpt-6-astra at max

Review point (g) of the Gate A owner: the sealed Gate A E2E runbook launches
Codex workers with `-m gpt-6-astra -c model_reasoning_effort="max"`
(evidence/artifacts/token-adoption-e2e-20260926/RUNBOOK.md:371; README.md:266),
so the handbook's `-m gpt-6.1-sol` command is marked as the post-D4 default
and names the sealed route beside it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: the Gate A owner's decisions (window-W guard, coordinator scoping), corrections, sources and numbers at final content

- DECISION 1 (record only): B1 applies neither user-level template;
  ~/.claude/CLAUDE.md (managed block) and ~/.codex/AGENTS.md stay
  byte-identical until the last Gate A window closes, F1's templates apply
  after window W. Why: the Codex block's line 16 names seven lane tools and
  every Codex arm reads the global AGENTS.md; the sealed design appends no
  LANES block (E2E README:188) and arm N is "config-free, not
  guidance-free" (README:275); Amendment 4 seals the pre-change hashes of
  both files. The guard covers window W only.
- DECISION 2: one wording on the three surfaces, coordinator-scoped
  imperatives, the pinned-launch rule (seed-binding-4, E2E README:157), the
  dropped "every manifest skill ..." clause (false at head) and why.
- Corrects the previous "makes no extra call" claim: the conditional
  wording gated only find-skills; search-first stayed ungated.
- Sources (point (f)): the pinned discovery command (skills-1.7.0 `skills
  find` with DISABLE_TELEMETRY=1, skills-agents-layer README:90,
  adoption/skills/manifest.json:22); the completeness critic cites
  Anthropic's evaluator-optimizer workflow and the multi-agent research
  system's completeness criterion; the north-star clause records "no
  upstream source; checked both pages, 2026-09-30".
- Numbers at final content (token_manifest count_files, gpt-tokenizer
  4.0.0, base main 1f2cdce): AGENTS.md 2,392 -> 2,812 (+420), portable
  2,051 -> 2,506 (+455), sum +875 against 200 (+723 accepted by the
  coordinator; +152 is the review's required text); Codex block 829 -> 1,397
  (+568); scaffold 373 -> 941. Words are Python str.split() counts.
- Post-A3 step marked done; six residual links listed (four F3, two D4).

docs/harness-defaults.md logs the proven mistake the review found: stating
a rule or a measurement for the head from an earlier state.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: attribute the Amendment 4 seal to the Gate A owner's review

The amendment text is not in this repository, so the record states the seal
on the two user-level files' pre-change hashes as the review's statement,
not as a fact verified here.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

---------

Co-authored-by: Scout <scout@local>
Co-authored-by: Claude Opus 5.5 <[email protected]>

…s; task-to-model routing record (unit A4) (#540)

* Codex user template: standing rule clauses, skills, models, routing and token lanes

Adds the 2026-09-30 standing clauses to the top-rule block of
adoption/templates/codex.AGENTS.template.md: skill discovery (search-first,
find-skills, skill-creator; implicit invocation from a skill's description and
explicit $skill-name, per openai/codex rust-v0.159.2
codex-rs/ext/skills/src/catalog_prompt.rs:8), upstream A/B and E2E harnesses,
the completeness critic feeding the next landscape sweep, the north-star
direction, GPT-6 model routing through the OmniRoute gateway (codex -p
omniroute), dated decision records, the startup rule, and a token-lanes line
naming each MCP server the Codex config template registers.

The RTK upstream text and the exceptions block stay byte-identical, so the F4
block in every Codex role is unchanged. The top-rule pin is re-derived with
template_segments(): 455 words, 7b41478f...; a new test requires a lane for
every server in codex.config.template.toml and the standing phrases.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Portable Claude template: superset of the user-level rules plus the standing clauses

examples/claude-native/CLAUDE.md becomes the single managed source of the
operator's user-level file. Step 1 of the top rule takes the user-level
paragraph (research and record, reference implementations, popularity guides
discovery) and the skill-discovery clause (search-first, find-skills with
`npx skills find`, skill-creator); the core rule takes the rules only the
user-level file held (short plan, pins and reasons, relevant layers, reviewer
agreement is not proof), the north-star direction, the upstream A/B and E2E
harnesses and the completeness critic; token practice takes the startup rule;
the worker section takes GPT-6 routing through the OmniRoute gateway. Every
worker, model, Ultracode and agent-team rule is kept word for word.

PortableTopRuleTests gains a phrase check for the clauses and the user-level
rules (it failed on the unedited template with 29 phrases missing, 8 of them
rules the user-level file held) and re-baselines the word budget to 1,656.
@RTK.md stays the host's own import (recipes/claude-native-profile.md:134).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Decision record: standing rule text in every instruction layer; divergence log row

docs/decisions/2026-09-30-rule-text-every-layer.md supersedes the 2026-09-28
top-rule record's "the operator's user-level file is the operator's own" at
the user's 2026-09-30 request: the portable template becomes the single
managed source of that file. It maps the six clauses to each surface, records
the alternatives (a UserPromptSubmit carrier, per the hooks reference, adds
its context beside every prompt; leaving the file unmanaged produced the
divergence; per-agent copies are unit F2's; pointers measured as the
fallback), the o200k sizes from token_manifest.count_files, the overrun of
the unit's 200-token budget with its floor and fallback, and the overturn
conditions.

docs/harness-defaults.md logs the host/template divergence, naming only the
check seen failing on it (the new PortableTopRuleTests phrase check).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* AGENTS.md: the six standing clauses on the four named lines; re-register hashes

Hot files last (docs/lanes.md hot-file protocol). AGENTS.md takes the
standing clauses by amending, not appending, the lines the brief names:
- line 3, the top rule: skill discovery (search-first, find-skills with
  `npx skills find`, skill-creator; manifest skills model-invocable in both
  clients) replaces "with the installed research and skill-discovery
  skills", and A/B and E2E on upstream harnesses, never a self-written runner;
- line 7: the north-star R&D direction; each unit names its action;
- line 18: the completeness critic feeding the layer's next landscape sweep;
- line 28: GPT-6 through the OmniRoute gateway (astra max, sol medium, 6.1
  sol pending qualification), Codex CLI as the second native client, and "no
  audits, trials or network at startup" with the daily currency timer's
  due-file line (record lands with unit A2) replacing "Do not rerun the full
  audit or model trials at startup".

manifests/evidence.json re-registers the six changed files already listed,
with scripts/host_receipts.py register_file; component_matrix and
new_host_grand_list --check pass unchanged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record and log row: nine user-level rules, not eight

The phrase check's committed form (with "from the selected source revision",
the user-level file's own wording) finds 29 phrases missing from the unedited
template, 9 of them held only by the user-level file; the record, its Checks
section and the anti-pattern row said 8 (from a run before that phrase
changed). The record also states the two non-verbatim worker lines exactly:
"Quality comes first" is one line with an identical word sequence, and the
sizing line keeps every user-level word and adds "to its task".
Re-registers docs/harness-defaults.md (hot file, last commit).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Codex user template: Sol-primary routing and skill matching folded into the standing lines

Folds the two lines the main checkout's uncommitted changes added inside the
top-rule block (pre-existing uncommitted changes observed in the main
checkout; original author not established), merged with this unit's lines
instead of carried beside them:
- the Models line now states the Sol-primary routing of
  docs/decisions/2026-09-30-sol-primary-quality-defaults.md (gpt-6.1-sol at
  ultra for coordination and at max for workers, gpt-6-astra at max after the
  recorded escalation triggers), replacing the superseded astra/sol-medium
  wording and "gpt-6.1-sol waits for qualification"; the OmniRoute gateway
  clause stays;
- the skills line takes the skill-matching rule (read each selected SKILL.md,
  follow its native workflow, load references only when needed) next to
  implicit and $skill-name invocation.

The phrase test failed first on the unedited block with five phrases missing
(exit 1). Pin re-derived with template_segments(): 496 words, 2d3107a2...;
the RTK text and exceptions block are unchanged (test_codex_agents 18 OK).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Portable Claude template: Sol-primary Codex routing; skill matching kept inside "Keep context small"

- The GPT-6 bullet follows docs/decisions/2026-09-30-sol-primary-quality-defaults.md:
  Codex CLI is the second native client, gpt-6.1-sol at ultra coordinates and
  at max runs workers, gpt-6-astra at max takes the recorded escalation
  triggers; cross-family research, review and sweep votes run through the
  OmniRoute gateway. It replaces the superseded astra/sol-medium wording and
  "gpt-6.1-sol pending qualification".
- Folds the skill-matching wording of the main checkout's uncommitted change
  (pre-existing uncommitted changes observed in the main checkout; original
  author not established) into the token-practice bullet. The change as
  observed replaced "Keep context small", a rule the operator's user-level
  file holds; the merged bullet keeps both.

The phrase check failed first on the unedited template with four phrases
missing (exit 1); it now also requires "Keep context small", which the
observed version lacks (mutant check: 31 phrases missing there, that one
included). Word baseline 1,656 -> 1,680.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* AGENTS.md: fold the skill-lifecycle and Codex-defaults bullets; one model-routing sentence set

Folds the two bullets of the main checkout's uncommitted AGENTS.md change
(pre-existing uncommitted changes observed in the main checkout; original
author not established). Neither was on origin/main at 11227bf (the file's
blob there equals e45328d's) nor on any other unit branch:
- the token-practice bullet on matching skill descriptions, reading each
  selected SKILL.md and adoption/skills/lifecycle.md (which unit F3 adds);
- the Codex defaults bullet (gpt-6.1-sol / ultra coordinator, gpt-6.1-sol /
  max primary workers, Astra/max on the recorded triggers, the Sol-primary
  record that unit D4 adds), merged with this unit's clause: Codex CLI is the
  second native client, and cross-family research, review and sweep votes run
  through the OmniRoute gateway. The startup bullet keeps only the startup
  clause, so the model routing is stated once.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Convergence guide and token handbook: Ultracode decoupled from effort; Sol worker command

Folds the true delta of two files from the main checkout's uncommitted
changes (pre-existing uncommitted changes observed in the main checkout;
original author not established), taken as a three-way merge onto
origin/main so main's own later edits stay:
- docs/convergence-architecture.md: Claude Code 2.1.284 decoupled Ultracode
  from effort; xhigh (saved fallback) and max (launcher) are separate choices,
  not measured quality gains. Source: the tagged anthropics/claude-code
  v2.1.285 CHANGELOG.md line 206 ("it no longer forces xhigh effort and stays
  on at any effort level", under 2.1.284), read 2026-09-30.
- docs/token-session-handbook.md: the Codex worker command names
  gpt-6.1-sol, matching the stack-worker profile and recipe that unit D4
  changes in the same batch.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Anti-pattern log: ten folded rows and six Codex runtime lane rows

Ten rows are folded byte-for-byte from the main checkout's uncommitted
docs/harness-defaults.md (pre-existing uncommitted changes observed in the
main checkout; original author not established), at the top of the table
where that change placed them. Only the true delta against origin/main is
taken: main's newer login-shell row and the four terminal-lane rows of #532
stay; the "Sol-primary quality defaults" paragraph belongs to unit D4.

Six rows were requested by the Codex runtime lane (relays
codex-runtime-anti-pattern-owner-handoff-20260930 and
codex-f1-source-correction-handoff-20260930), one mistake / correction /
check each, citing that lane's published sources at full SHAs: PR #535 head
00aa6c2 (the cited files are byte-identical to the earlier head 6a7b164,
which no remote ref holds any more), SDK branch head 404b821 and OpenHands
software-agent-sdk dcf401af build.py lines 581 and 925. The GitHub Actions
job conclusions the rows cite were re-read through the REST jobs API on
2026-09-30 (run 36695388851: jobs 109821999811 and 109821999568 cancelled,
all steps success; run 36690153586: job 109805172026 validate-macos failure).

UpstreamVerificationSectionTests: 3 OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Fold the 2026-09-30 native practice finalization record, attributed

docs/decisions/2026-09-30-sota-native-finalization.md is folded unchanged
from a read-only snapshot of the main checkout (pre-existing uncommitted
changes observed in the main checkout; original author not established;
tracked diff sha256 314bd1b260da0939). One attribution paragraph after the
title states that provenance and which linked evidence is not published with
it: evidence/artifacts/sota-finalization-20260930/ has no assigned owner, the
native-skill-finalization artifacts land with unit F3 and the
codex-01592-qualification artifacts wait on unit D4. The body is
byte-identical to the snapshot (12,956 bytes); the privacy scan found no host
path, home directory or host user name in it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: rebuilt base, Sol-primary clause (a), folded items, lane rows and guards, re-measured sizes

- Base moves to origin/main@11227bfd (the three surfaces are byte-identical
  to e45328d's).
- Clause (a) follows docs/decisions/2026-09-30-sol-primary-quality-defaults.md
  (unit D4); the brief's first wording and its 04:35Z "Astra at ultra for
  complex workflow tasks" relay are recorded as superseded by that later,
  more specific user selection.
- Lists every item folded from the main checkout with the provenance
  sentence, the snapshot hashes, the three-way-merge method and what was not
  folded (recipes/README.md is on D4's branch; the finalization evidence
  directory has no owner).
- Records the six Codex runtime lane rows (relays and what was verified) and
  the three relayed guards: #543 not described as published without a native
  gh lookup that exits 0, the 07:34Z/06:46Z actor unknown, a historical OSV
  pass not current after #546.
- Sizes re-measured with tools/token-report/token_manifest.py count_files
  (gpt-tokenizer 4.0.0): AGENTS.md +323, portable template +363 (sum +686
  against the 200 budget: +139 folded text, +547 clauses and user-level
  rules), Codex template +507.
- Limitations updated: clause (c) against F3's settings, cross-unit
  references, unpublished evidence, older Codex pins, A3's scaffold copy.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Codex routing on all three surfaces: Astra/Ultra coordinates a complex workflow, Astra/Max takes a single consequential judgment

Follows the coordinator's decision and unit D4's refined Sol-primary record
(claude/sota-defaults-d4-codex-0159-20260930 at 9dd4ebe), which splits the
escalation by the Codex catalog semantics it documents (Ultra: proactive
delegation with the model's xhigh reasoning; Max: the highest reasoning
effort) and quotes the user's 2026-09-30 selection ("astra ultra when tasks
needed suitable for complex workflow"):
- gpt-6-astra at ultra when a complex workflow needs Astra to coordinate it;
- gpt-6-astra at max for a single consequential judgment (conflicting primary
  evidence, consequential architecture, complex changes across systems, or a
  failure unresolved after one bounded Sol repair).

AGENTS.md's folded Codex bullet, the portable template's Codex bullet and the
Codex block's Models line now state the split. Both phrase checks failed
first with four phrases missing each (exit 1). Codex pin re-derived with
template_segments(): 515 words, d1195686...; portable word baseline 1,697.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* One-pass tightening of the added rule text (no rule removed)

The coordinator accepted the growth over the 200-token budget and asked for
one pass that cuts words carrying no rule, limited to text this branch adds:
- AGENTS.md Codex bullet: the Astra split as two clauses and the contract
  pointer in parentheses (2,742 -> 2,733 o200k tokens);
- portable template: "(registry: `npx skills find`)" -> "(`npx skills find`)"
  (2,436 -> 2,433);
- Codex block: implicit/explicit invocation in one clause, "use the
  `search-first` skill" -> "use `search-first`", the registry label, and
  "the paired benchmark of Claude's `skill-creator` plugin" -> "Claude's
  `skill-creator` paired benchmark" (1,360 -> 1,346).

Every phrase of both phrase checks still holds. Codex pin re-derived with
template_segments(): 507 words, 97bbeb8c...; portable word baseline 1,696.
Counts: tools/token-report/token_manifest.py count_files, gpt-tokenizer 4.0.0.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Fold the finalization record's evidence: 18 of the 19 files of evidence/artifacts/sota-finalization-20260930/, unchanged

Allowed-path amendment by the coordinator. The files are copied
byte-for-byte from the read-only snapshot r2 of the main checkout
(pre-existing uncommitted changes observed in the main checkout; original
author not established); sha256sum -c against the snapshot passes.

Sanitization: nothing needed stripping and no file was changed.
- validate.py --scan-file over all 19 snapshot files: {"scanned_files": 19,
  "status": "passed"} (UUID, personal home path, Windows user path and token
  patterns).
- A wider scan found no personal home path, host user name, /tmp/claude-*
  path, session UUID, e-mail, IP address or token. The one "/home/" string
  earlier reported as a home path is the literal placeholder "/home/example"
  inside claude-review.json's prose, which validate.py's pattern exempts.
  Replacing it with <home> would alter the retained review without removing
  any host data, and would break the sha256 pin that the lane's convergence
  record holds for that file, so the file stays byte-exact. The
  config-key hits in convergence.json are recorded codex exec command lines
  already redacted to <private-path>; the three "~/." strings are generic
  install locations.

Held back: convergence.json. It is a kind "convergence_experiment" record
whose frozen inputs and observations pin five files of
native-skill-finalization-20260930 (unit F3; all five pins match F3's branch)
and three of codex-01592-qualification-20260930 (on no branch yet) by path and
sha256. validate_convergence.py --all-recorded, which CI runs, fails
discovery for an undeclared hash-listed record and fails validation for a
declared one with missing files, so it can be folded, declared in
convergence_records, only after F3 lands and D4 publishes those three files
unchanged at those paths.

The folded record's attribution paragraph says so; its links and the four
folded log rows into this directory resolve. Hash registration follows in
the last commit.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: Ultra/Max split, accepted budget miss with its overturn condition, evidence fold and sanitization, post-A3 step

- Base is origin/main@8fc86119 (rule surfaces byte-identical to 11227bf
  and e45328d).
- Clause (a) records the split from unit D4's refined record (9dd4ebe):
  gpt-6-astra at ultra when a complex workflow needs Astra to coordinate it,
  at max for a single consequential judgment, with the catalog semantics and
  the user's 2026-09-30 words as relayed.
- Measured size: AGENTS.md 2,392 -> 2,733 (+341), portable template
  2,051 -> 2,433 (+382), Codex block 829 -> 1,346 (+517); budget sum +723
  against 200, accepted by the coordinator because the clauses are the user's
  explicit directive. Composition: fold +139, split +49, tightening -12.
  Overturn condition: a rerun of the 2026-09-29 user-prefix A/B showing no
  rule-following gain for the added cost brings back the pointer fallback.
- The finalization record's evidence: 18 of 19 files folded unchanged; the
  sanitization section gives the scan and corrects the first handoff (19
  files, not 21; the one "/home/" string is the /home/example placeholder,
  kept byte-exact, also because the lane's convergence record pins its
  sha256). convergence.json is held back: a convergence_experiment record
  pinning F3 and D4 files that CI's validate_convergence --all-recorded
  would reject on this branch whether declared or not.
- Known residual links: seven targets, four from F3 and three from D4
  (the Sol-primary record and two codex-01592-qualification files).
- Post-A3 step: after #545 merges, whichever lands second copies the
  top-rule block into adoption/scaffold/AGENTS.md; not created here.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule text: conditional skill discovery on the three surfaces; the growth waiver and the provenance wording cited to the brief's amendments

find-skills (npx skills find) runs only when no listed skill fits the task, so a measured child with a fitting
listed skill makes no extra Skill or Bash call (the Gate A owner's note); search-first stays before custom code
or a tool choice. The record cites the coordinator's dated brief amendments for the accepted +723 o200k growth and
for the neutral fold provenance (a process census of file writes does not establish authorship; the Codex runtime
lane asked for the wording). Codex top-rule pin re-derived: 514 words, fecf92dc...; portable baseline 1,703 words.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

* Standing rule text: one wording on the three surfaces, coordinator-scoped (Gate A owner's review of #557)

Applies DECISION 2 of the Gate A owner's ACCEPT-WITH-CHANGES review on
AGENTS.md, examples/claude-native/CLAUDE.md and the Codex block, with the same
sentences on all three (the Codex block names a bounded worker where the
Claude surfaces name a delegated child in the skill sentence):
(a) "A coordinator, not a delegated child, invokes `search-first` before
    custom code or a tool choice; when no listed skill fits the task, it
    discovers one with `find-skills` and verifies or A/B-tests it with
    `skill-creator`."
(b) Sol/Astra routing kept for unpinned work; added: where a launch pins
    the model and effort (`-m`, `-c model_reasoning_effort`), children inherit
    that pin and a spawn call names neither; a coordinator, never a
    delegated child, starts a cross-family lane.
(c) Dropped "every manifest skill stays listed for model invocation in both
    clients" (false at this head: find-skills, grill-me and
    improve-codebase-architecture are user-invocable-only, find-skills has
    codex_enabled false) and the bare `npx skills find`.
(d) The north-star, completeness-critic, trigger/acceptance and dated
    decision-record imperatives are a coordinator's.
(f) AGENTS.md's lifecycle pointer says "(lands with unit F3)"; the
    currency-notice (A2, in the stack below) and Sol-primary (D4, on main)
    citations stay.
The harness and startup sentences take one wording too.

Post-A3 step done: adoption/scaffold/AGENTS.md carries the final top-rule
block byte for byte (tests.test_scaffold_repo passes).

Tests: StandingRuleSurfacesTests (new) asserts the shared sentences on all
three surfaces and the dropped clause on none; it failed first with 21
failing subtests (exit 1). Both phrase lists failed first (exit 1). Codex pin
re-derived with template_segments(): 538 words by Python str.split(),
9565217f...; portable baseline 1,750; comments now name str.split(), not
wc -w (review point (e)).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token handbook: the Sol worker command is marked "after D4"; the sealed E2E route keeps gpt-6-astra at max

Review point (g) of the Gate A owner: the sealed Gate A E2E runbook launches
Codex workers with `-m gpt-6-astra -c model_reasoning_effort="max"`
(evidence/artifacts/token-adoption-e2e-20260926/RUNBOOK.md:371; README.md:266),
so the handbook's `-m gpt-6.1-sol` command is marked as the post-D4 default
and names the sealed route beside it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: the Gate A owner's decisions (window-W guard, coordinator scoping), corrections, sources and numbers at final content

- DECISION 1 (record only): B1 applies neither user-level template;
  ~/.claude/CLAUDE.md (managed block) and ~/.codex/AGENTS.md stay
  byte-identical until the last Gate A window closes, F1's templates apply
  after window W. Why: the Codex block's line 16 names seven lane tools and
  every Codex arm reads the global AGENTS.md; the sealed design appends no
  LANES block (E2E README:188) and arm N is "config-free, not
  guidance-free" (README:275); Amendment 4 seals the pre-change hashes of
  both files. The guard covers window W only.
- DECISION 2: one wording on the three surfaces, coordinator-scoped
  imperatives, the pinned-launch rule (seed-binding-4, E2E README:157), the
  dropped "every manifest skill ..." clause (false at head) and why.
- Corrects the previous "makes no extra call" claim: the conditional
  wording gated only find-skills; search-first stayed ungated.
- Sources (point (f)): the pinned discovery command (skills-1.7.0 `skills
  find` with DISABLE_TELEMETRY=1, skills-agents-layer README:90,
  adoption/skills/manifest.json:22); the completeness critic cites
  Anthropic's evaluator-optimizer workflow and the multi-agent research
  system's completeness criterion; the north-star clause records "no
  upstream source; checked both pages, 2026-09-30".
- Numbers at final content (token_manifest count_files, gpt-tokenizer
  4.0.0, base main 1f2cdce): AGENTS.md 2,392 -> 2,812 (+420), portable
  2,051 -> 2,506 (+455), sum +875 against 200 (+723 accepted by the
  coordinator; +152 is the review's required text); Codex block 829 -> 1,397
  (+568); scaffold 373 -> 941. Words are Python str.split() counts.
- Post-A3 step marked done; six residual links listed (four F3, two D4).

docs/harness-defaults.md logs the proven mistake the review found: stating
a rule or a measurement for the head from an earlier state.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Rule-text record: attribute the Amendment 4 seal to the Gate A owner's review

The amendment text is not in this repository, so the record states the seal
on the two user-level files' pre-change hashes as the review's statement,
not as a fact verified here.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Task-to-model routing record: one table of task class, client, model, effort and enforcement point

Consolidates the routing rules of the workflows README (dispatch by role, Sonnet 5.5
fan-out units, role routing), the model-currency record, the Sonnet 5.5 dispatch
record and the max-default effort record into one table, and names where each route
is enforced today: agent frontmatter, workflow stage, CLAUDE_CODE_SUBAGENT_MODEL,
client settings, Codex config, lane code or instruction only. No route changes.

States that no automatic router exists; OmniRoute D04 singleton combos, D06 lane
aliases and the task-aware router (D11) stay off until the promptfoo A/B of wave
unit D3. gpt-6.1-sol is listed only as pending unit D4 qualification. Open PRs #423
and #508 are cited at their head commits.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* token-practice: point to the task-to-model routing record

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Test the routing record against the enforcement points it quotes

Every value the record's table quotes from a file must still be on the cited
lines, every cited line must exist, frontmatter rows must name the model they
quote, every task class must have a row, gpt-6.1-sol must stay listed only as
pending and routed nowhere, and docs/token-practice.md must point to the record.
Integration checks of the repository's own record, not upstream acceptance.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Routing record: name the client's content-based fallback guards

"No automatic router exists" now also names Claude Code's own content-based fallback
and the two keys that switch it off in the project settings and the user-settings
template (switchModelsOnFlag false, CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1), with the
model-currency addendum's note that their coverage is unverified and that availability
fallback chains are a separate mechanism.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Routing record: cite #508 line 90 alone for the run-shape levers

#508's token-stack record says the run-shape levers, role dispatch
among them, are "owned by the Gate A owner after #381 closes" (line 90).
Its line 93 ("Gate A per-row results decide any change") is about the
on-demand members of line 92, not the levers, so the Context paragraph
and the Sources entry now cite line 90 only (review finding, low).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Routing record and test: hold quoted values per file, not per line; Codex rows name no version

Six sibling units of this wave may edit files that the record quotes by
line: landscape-sweep (A1), bootstrap-linux.sh (A3), .claude/agents and
the settings template's hooks entry (F2), the settings template and the
Codex template (F3), codex_roles.py (F4) and the model-currency record
(D4, an addendum over the region of its line 284). None of them may edit
this record or its test. An exact-line check fails their pull requests
on any inserted line, even when no route changes.

- The record now states every line number as read at e45328d.
  tests/test_task_model_routing.py counts each quoted value in its whole
  file and requires at least one copy for each distinct line quoted, so
  a pure line shift passes and a changed model or effort fails.
- Two values have more copies than the record cited. It now also cites
  the template's modelSettings xhigh for claude-opus-5-5 and
  claude-sonnet-5-5 (:303, :306) and review-changes.js's recheck-stage
  scout binding (:78), so a change to any cited copy fails.
- model-currency.md:284 is cited without a CI-checked quote. D4 may
  rewrite that sentence.
- The Codex rows' Client cell says "Codex CLI". The preamble records that
  manifests/stack.json:297 pins 0.157.1 at e45328d, so D4's pin move
  leaves no stale version string (review finding, low).
- The GPT-6.1 check matches a model binding (a model key or constant, or
  -m/--model with an optional gateway prefix), not any mention. F1 may
  write prose naming gpt-6.1-sol into adoption/templates; a real binding,
  such as D4 switching the Codex template default, still fails.

Controls, each run in a fresh copy of the files the test reads:
16 of 16 as expected for both the old and the new test. Pure line
shifts (codex template, model-currency.md, sweep.js) and an F1-style
prose mention fail the old test and pass the new one. Changed frontmatter,
one of two identical sweep bindings, and a gpt-6.1-sol template default
fail both. Changing the :306 per-model xhigh or the :78 recheck binding
passes the old test and fails the new one.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* token-efficiency profile: accept the selection, carry the three code-navigation tools

The profile's label drops "Drafted, not accepted" for a dated acceptance of the
selection, citing the routing record, which joins recipe_paths so the profile
resolves only while the record is present. jcodemunch-mcp, codebase-memory-mcp
and ast-grep join component_ids and required_commands: the SubagentStart carrier
names them as task-appended lanes and the code-navigation layer's current choice
names all three (catalogs/landscape/foundation.json). Their versions stay in
manifests/stack.json (1.108.319, 0.11.0, 0.45.3); a manifest profile has five
keys and carries no version or wiring field, so neither is added.

tests/test_adoption_status.py restates the split the profile is checked against:
NAVIGATION_CHOICE, asserted against the code-navigation current choice, leaves
OPTIONAL, and the Ultracode tool split becomes 13 profile rows and 3 optional
rows against 17 component_ids.

Failing-first: the restated tests fail against the previous manifest (2 of 3 in
TokenEfficiencyProfileTests) and the previous tests fail against this manifest
(the same 2, plus ProfileTableTests.test_pin_columns_match_the_pin_files, whose
adoption/README.md cells are outside this unit's paths). manifests/evidence.json
is re-registered in the branch's last commit.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* Routing record: accept the token-efficiency profile, state the three tools' wiring

The record no longer says the profile is "not decided here"; it makes the claim
that adoption/manifest.json now cites. The Decision gains a dated acceptance that
bounds itself to the selection and its routing (structural validation, not a host's
acceptance), the pin and the Claude Code and Codex wiring of jcodemunch-mcp,
codebase-memory-mcp and ast-grep with the file that carries each, the basis for
adding them (the SubagentStart carrier, the code-navigation current choice, the
2026-09-25 Ultracode run), and where #508 differs: it lists ast-grep as the
structural-code lane, makes jCodeMunch an owner only once wired and lists
codebase-memory-mcp as not a member, so the record does not rest on #508 for those
two. Two alternatives (leave the three optional; a new manifest key) and two
overturn conditions (#508 or Gate A changes a row; pins arrive) are added, and the
Sources name the new files and #508 lines 85, 86 and 103.

Every path:line cite in the whole record resolves in this branch (115 cites, 28
bare paths); the record keeps its five sections and one table, which
tests/test_task_model_routing.py holds.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* token-efficiency-stack: restate the optional-row split for the three tools now in the profile

Three statements had become false: that jCodeMunch, ast-grep and codebase-memory-mcp
are optional rows outside the profile, that the profile has 14 component_ids, and
that the Ultracode run's 16 tools are ten profile rows and six optional rows. The
profile paragraph now lists the three as the code-navigation layer's task-selected
tools (none a layer winner), the optional-row paragraph keeps Context Hub, the
viewers and OmniRoute, and adds why client_wiring still checks only Serena,
SocratiCode and ai-memory, that the three have no platform pin (so --pinned-versions
reports them unchecked) and that a bootstrap refuses the profile until they are
named in --allow-unpinned (adoption/bootstrap-linux.sh --help: exit 3). The
Ultracode paragraph becomes 17 component_ids, thirteen profile rows and three
optional rows, matching tests/test_adoption_status.py.

The review of this unit asked for this reconciliation (docs/token-efficiency-stack.md
lines 264-270); the file is outside the brief's allowed paths, so it is its own commit.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* Test the token-efficiency profile against the routing record that accepts it

The record's Decision says the profile's label accepts it and names the record,
that the record is one of its recipe_paths, and that jcodemunch-mcp,
codebase-memory-mcp and ast-grep are component_ids and required_commands. The new
test in tests/test_task_model_routing.py holds those claims to adoption/manifest.json
and requires the record to name each tool. It compares no version: manifests/stack.json
owns them, and the record now dates the ones it quotes at e45328d.

Failing-first: against the previous manifest the test fails on the label ('Accepted'
not found in 'Drafted, not accepted: ...'); against this branch's manifest all 8
tests in the module pass.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* token-efficiency profile: README pin cells, regenerated grand list and re-registration (companion)

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* token-efficiency profile: back to the 14 pinned components; the three code-navigation tools stay optional rows

Round 1 (24c8703) added jcodemunch-mcp, codebase-memory-mcp and ast-grep to the
profile's component_ids and required_commands. Neither adoption/pins-linux-x86_64.json
nor adoption/pins-macos-arm64.json lists them, and both bootstraps fail closed on a
selected component with no pin: adoption/bootstrap-linux.sh:169 and
adoption/bootstrap-macos.sh:220 print "No pin in <file> for selected component(s)" and
exit 3, and adoption/bootstrap-macos.sh:176-190 records that no component is exempted
from a pin by default. PR #540's validate-macos job failed on it (3 tests of
tests/test_adoption_bootstrap_macos.py TokenEfficiencyPlanTests), reproduced locally at
8ca7895 with `python3 -m unittest tests.test_adoption_bootstrap_macos
tests.test_adoption_bootstrap tests.test_adoption_launchd`: FAILED (failures=3,
skipped=38).

The profile is again the 14 pinned components. It keeps the accepted label and the
routing record in recipe_paths. The three tools stay the code-navigation layer's
task-selected tools and the carrier's task-appended lanes, installed on demand from their
recipes, until each has a reviewed pin on both platforms and a host receipt.

tests/test_adoption_status.py returns to main's OPTIONAL set of eight and the
(16, 14, 10) Ultracode split, and TokenEfficiencyProfileTests gains
test_every_profile_component_has_a_pin_on_both_platforms, so a profile that outruns the
pin files fails in this file's own module set and not only in the bootstrap plan tests.
tests/test_task_model_routing.py holds the record's claims about the three: outside
component_ids and required_commands, unpinned on both platforms, named by the
code-navigation current choice and by the carrier's lane ids.

Failing-first: with these tests and the round-1 manifest (17 component_ids), 4 of the 12
tests in TokenEfficiencyProfileTests and TaskModelRoutingRecordTests fail, 8 failures
counting subtests ("Tuples differ: (16, 17, 13) != (16, 14, 10)", "'jcodemunch-mcp'
unexpectedly found in [...]", the pin-file membership checks). With the manifest of this
commit all 12 pass.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* token-efficiency-stack: restore the optional-row split, since the three tools are not profile members

This reverts 12a587e. docs/token-efficiency-stack.md is byte-identical to origin/main
(11227bf) again: the profile has 14 component_ids, jCodeMunch, ast-grep and
codebase-memory-mcp are task-selected options in the code-navigation layer's current
choice and stay optional rows, and the Ultracode run's 16 tools are ten profile rows and
six optional rows. The statements 12a587e rewrote (three joined the profile on
2026-09-30; a bootstrap refuses the profile until they are named in --allow-unpinned; 17
component_ids, thirteen profile rows) described a profile that adoption/manifest.json no
longer has, because both bootstraps fail closed on a component without a pin and none of
the three has one. tests/test_adoption_status.py checks the ten/six split against the
receipt's tool list. The reasoning is recorded once, in
docs/decisions/2026-09-30-task-model-routing.md.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* adoption README: the token-efficiency row says accepted, and its cells describe the 14 pinned components again

d5369cb wrote the row for 17 members: "14 of 17 (all 14 at v2026.09.26.2)" in both pin
columns, the three code-navigation tools in the description, and a note that a
bootstrap of the profile needs them in --allow-unpinned. The profile has its 14 pinned
components again (previous commit), so the description, both pin cells and the note are
main's, and only the acceptance wording stays: "Accepted 2026-09-30 as the selection",
linking the routing record.

The pinned release v2026.09.26.2 covers the same 14 of 14 as HEAD, so
tests/test_adoption_docs_consistency.py (ProfileTableTests.cell_errors) asks for no
note on that tag; the notes for v2026.09.25.2 and v2026.09.26 are checked against
their own tags and hold.

Failing-first: with the round-1 row and the 14-component manifest,
ProfileTableTests.test_pin_columns_match_the_pin_files fails ("token-efficiency Linux
pins: table says '14 of 17', adoption/pins-linux-x86_64.json gives 'all 14'"). With
this row the module passes: Ran 37 tests, OK (skipped=1). The one skip is data
dependent, not environmental: no profile's coverage differs between the pinned release
and HEAD, on origin/main as well as here.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* Routing record: the three code-navigation tools stay outside the accepted profile, for want of a pin

The record accepted the token-efficiency profile and, in round 1, added jCodeMunch,
codebase-memory-mcp and ast-grep to it. It now accepts the profile as it is, its 14
pinned components, and says why the three are not members: both bootstraps fail closed
on a selected component with no pin (adoption/bootstrap-linux.sh:144-172,
adoption/bootstrap-macos.sh:177-223, and the macOS comment at :177-187 that no component
is exempted from a pin by default any more), none of the three has an entry in
adoption/pins-linux-x86_64.json or adoption/pins-macos-arm64.json, and
tests/test_adoption_bootstrap_macos.py:1015-1017 requires the profile's plan to resolve
every component with no --allow-unpinned.

- Alternatives: leaving the three out is the chosen state; adding them (round 1) is a
  rejected alternative with its failure (Adoption bootstrap smoke run 36728291629, job
  validate-macos, step "Gate on the adoption test modules"; reproduced locally at
  8ca7895, failures=3) and the reason --allow-unpinned does not rescue it.
- Decision: "Three tools join the profile" becomes "Three tools stay outside the
  profile"; the wiring bullets, which describe how each installs on demand, stay; the
  "14 of 17" coverage and the exit-3 sentence are gone; "Where #508 differs" becomes
  "#508's rows": leaving the three out agrees with #508 for jCodeMunch (line 86) and
  codebase-memory-mcp (line 103), and ast-grep (line 85, the structural-code lane) is
  out only for want of a pin.
- Overturn condition: the follow-up unit that adds a tool once it has a reviewed pin on
  both platforms and a host receipt for each, with the files it must restate in the same
  change (adoption/README.md row, token-efficiency-stack.md split, OPTIONAL and
  CARRIER_TOOLS in the two tests).
- Header and Sources: the citations are at e45328d while the branch is based on
  11227bf, where examples/claude-native/workflows/README.md has grown by 249 lines; the
  record says so. The pin-rule files it now cites are unchanged between the two bases.

tests/test_task_model_routing.py (previous commit) holds these claims to the files.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* Test the routing record against D4's Codex routes

tests/test_task_model_routing.py predated D4 (#542, merged as 1f2cdce) and failed on the
stacked tree (validate job 110072600341): two quoted `model = "gpt-6-astra"` lines and the
test that expected GPT-6.1 Sol to be routed nowhere.

- A quoted value counts only as a whole value, so `model = "${CODEX_MODEL}"` is not met by
  the template's `default_subagent_model` line.
- GPT-6.1 Sol: the table's Sol rows are exactly the Sol-primary record's three routes (the
  Codex coordinator at ultra, primary workers and generic children at max), and every file
  in the routing places (now with recipes/ and adoption/agents/) that binds a GPT-6.1 model,
  a constant such as CODEX_MODEL_CURRENT included, is cited by a Sol row. __pycache__ and
  node_modules are skipped: CI flagged prove_codex_lane's bytecode.
- The Codex user template is rendered through tools/adoption/render_config.py for every
  pinned platform, as tests/test_render_config.py's CodexModelTests do, and must give the
  models and efforts that the coordinator and generic-children rows name.
- The task classes add complex-workflow coordination and a single consequential judgment.

Fails against the record at this commit (5 failures, 1 error); the next commit restates it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Routing record: restate the Codex rows for D4 (#542)

D4 made GPT-6.1 Sol the Codex coordinator (ultra) and primary-worker model (max), with GPT-6
Astra for complex-workflow coordination (ultra) and a single consequential judgment (max):
docs/decisions/2026-09-30-sol-primary-quality-defaults.md and the model-currency addendum of
2026-09-30. The record still listed Sol as pending and routed nowhere.

- Six Codex rows, read at 1f2cdce: the Codex judgment role carriers (Astra, max); the
  coordinator and interactive default (the template's ${CODEX_MODEL} at ultra, rendered as
  gpt-6.1-sol for the Linux pin 0.159.2 and gpt-6-astra for the macOS pin 0.155.1); primary
  workers (the stack-worker profile, the lane's worker pins and live proof, the recipe's
  command); generic children; complex-workflow coordination and escalation (instruction only).
- Context and Decision state the D4 routing and its dependence on the Codex pin. No
  automatic router exists, and D04/D06/D11 stay off until the D3 promptfoo A/B, as before.
- Overturn: the user's next decision or D4's reopening condition, and a Codex pin crossing
  0.159.1 on a platform. Sources: D4's records, the enforcement points and the CI job.

The Claude rows are unchanged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Test each GPT-6.1 binding against the Sol rows, by TOML table and command profile

The GPT-6 cross-family review of round 2 (901ed76, needs_changes, medium): the "nowhere
else" check compared files only, so a `[profiles.security-review]` table with
`model = "gpt-6.1-sol"` added to the user template, a file the coordinator row already
cites, left all nine routing tests passing.

- Bindings are counted one by one: a TOML key by its table (parsed with tomllib: `model`,
  `agents.default_subagent_model`, `profiles.<name>.model`, on the ID or on
  `${CODEX_MODEL}`), a key or constant elsewhere (CODEX_MODEL_CURRENT), and a command's
  `-m`, named with the profile the command selects (`-p stack-worker -m`).
- Each binding must be one a Sol row cites by file and name. A quoted TOML key names the one
  key of that name and value (a quoted `[table]` header scopes it), and a quote that fits
  two keys is reported. Each cited binding must be one D4 makes: the template's `model` and
  `[agents] default_subagent_model`, the render constant, the stack-worker profile's
  `model` and the stack-worker command's `-m`.
- Regression: the planted profile, in memory, on the ID and on the placeholder, and then
  cited by its table from the coordinator row; each variant is reported.

Fails against the record at this commit: the stack-worker command's `-m` in the profile's
header (line 3) and in the live proof's help (tools/adoption/prove_codex_lane.py:36) are
bindings no Sol row cites. The next commit cites them.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Routing record: cite the stack-worker command where the profile header and the live proof write it

The binding-level check of the previous commit found two GPT-6.1 bindings that the
primary-worker row did not cite, both the stack-worker command's `-m`, which D4 routes: the
profile's header (adoption/templates/codex.stack-worker.config.toml:3) and the live proof's
help (tools/adoption/prove_codex_lane.py:36). The row now quotes both, beside the recipe's
command it already quoted. No other row changes.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Re-register the changed hash-listed files in manifests/evidence.json (hot-file protocol)

The last commit of the branch carries the hot files: the stacking ref's manifests/evidence.json
with this branch's changed hash-listed files re-registered, and the regenerated component-evidence
matrix and new-host grand list (docs/lanes.md hot-file protocol).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

---------

Co-authored-by: Scout <scout@local>
Co-authored-by: Claude Opus 5.5 <[email protected]>

…signal, spawn guard observed, sweep coverage stated, tmux identity control)

One read-only Claude Opus agent read the merged signal handling and shutdown code of the terminal lane (#532, #552) and returned 11 findings (3 medium, 8 low, none high). Each was verified by a measurement or by reading before it was repaired:

- the pty probes now end BY a latched signal (default action restored, then the signal sent to themselves) instead of exiting 128+N: a bash loop goes on after an exit 130 and stops after a death by SIGINT (shell_loop_control.py), as Cracauer's note says
- the per-line sweep records whether a client was started before the spawn guard and fails when one was; it states the lines swept of the lines with code, and the documents say what it does not reach
- an inherited SIG_IGN stays ignored; typing stops at a latched signal
- the tmux probes get a third shutdown-failure control that only the command-line check of the identity can fail (tmux_identity_mutant.py), a zombie no longer counts as alive, the pty.fork child ends with os._exit(127)
- timeouts count as failed results and still kill the stand-in; the mutant runner has eleven mutants and eight cases per probe (12 runs each, 0 problems)
- record of the read, rebuilt receipt, recorded outputs of 35 commands (Claude Code 2.1.286 scanned: the same 17 notification types), decision record, anti-pattern rows

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…signal, spawn guard observed, sweep coverage stated, tmux identity control) (#563)

One read-only Claude Opus agent read the merged signal handling and shutdown code of the terminal lane (#532, #552) and returned 11 findings (3 medium, 8 low, none high). Each was verified by a measurement or by reading before it was repaired:

- the pty probes now end BY a latched signal (default action restored, then the signal sent to themselves) instead of exiting 128+N: a bash loop goes on after an exit 130 and stops after a death by SIGINT (shell_loop_control.py), as Cracauer's note says
- the per-line sweep records whether a client was started before the spawn guard and fails when one was; it states the lines swept of the lines with code, and the documents say what it does not reach
- an inherited SIG_IGN stays ignored; typing stops at a latched signal
- the tmux probes get a third shutdown-failure control that only the command-line check of the identity can fail (tmux_identity_mutant.py), a zombie no longer counts as alive, the pty.fork child ends with os._exit(127)
- timeouts count as failed results and still kill the stand-in; the mutant runner has eleven mutants and eight cases per probe (12 runs each, 0 problems)
- record of the read, rebuilt receipt, recorded outputs of 35 commands (Claude Code 2.1.286 scanned: the same 17 notification types), decision record, anti-pattern rows

Co-authored-by: Scout <scout@local>
Co-authored-by: Claude Sonnet 5.5 <[email protected]>