Fix #1641: default the claude consult lane to the 1M context window

[mohidmakhdoomi]

Fixes #1641

The bug

consult -m claude failed repeatedly on PR #1640 (41 files, +4877/−110). Five attempts
across two independent callers — one architect-side (--type integration), four
builder-side (--protocol pir --type impl) — produced no claude review file, leaving
porch next 1481 blocked on Run remaining consultations (claude).

Four of those five ended in Prompt is too long; the third builder-side attempt hit a
separate subscription usage limit and is not evidence of this bug. Of the other lanes run
against the same target, codex completed (239.4s, REQUEST_CHANGES); gemini did
not review — its artifact is a non-blocking skip notice (Gemini lane skipped — agy exited with code 1), so it is not evidence that the target was reviewable either.

Root cause

The claude lane shipped the bare model id claude-opus-5.

The Agent SDK's bundled Claude Code runtime derives its context budget from the model id.
Its window function has several branches — an env override, the [1m] suffix, a
beta-header branch, and a sonnet-4-6-gated branch — before falling through to a 200,000
default, so this is not a universal "1M with a suffix, 200K otherwise" rule. The scoped
fact that matters here, verified on both runtimes this repo runs against — SDK 0.2.105
(the lockfile version, which bundles the runtime as cli.js) and 0.2.141 (the globally
installed build, whose JS package is a thin client over a separately packaged native
runtime). Both layouts are SDK-distributed and version-locked: 0.2.141 pins every
@anthropic-ai/claude-agent-sdk-<platform> optional dependency to exactly 0.2.141, and
SDK resolution locates that package's binary — it is not an arbitrary system CLI.

Model id	Budget
claude-opus-5	200,000
claude-opus-5[1m]	1,000,000

runClaudeConsultation runs the lane agentically — allowedTools: ['Read','Glob','Grep'],
maxTurns: 200 — so a review of a PR this size accumulates review context until the 200K
budget is exhausted, which surfaces as Prompt is too long. The two saved failing sessions
show substantial prior tool activity before the error: 29 and 51 tool calls, with
last-successful inputs of 174,521 and 175,930 tokens. That establishes accumulated-context
exhaustion; it does not establish where exactly the request was refused, nor that every
changed file was read in full. It is not an oversized opening prompt — the opening prompt
for issue 1481 inlines a 2.5K spec and a 34K plan, roughly 10K tokens.

The second half of the bug is why it was unrecoverable: MODEL_ID_RE rejected [ and
], so neither consult.models.claude nor --model-id claude-opus-5[1m] could work
around it.

The fix

• DEFAULT_CLAUDE_MODEL → claude-opus-5[1m].
• MODEL_ID_RE accepts one optional, case-insensitive [1m] suffix. The 1–200 character
  bound now applies to the id as a whole, via a lookahead, so the suffix counts toward
  the limit rather than escaping it.
• Validator error message updated to describe the suffix.
• Consult docs updated and kept byte-identical across codev/ and codev-skeleton/.

Explicitly configured model ids still pass through unchanged and --model-id still takes
precedence over config, so anyone pinning a bare id keeps today's behavior.

Opt-out. Setting consult.models.claude: "claude-opus-5" retains the previous
selection and its 200,000-token budget on the inspected runtimes. --model-id claude-opus-5 does the same for a single invocation, and
CLAUDE_CODE_DISABLE_1M_CONTEXT=1 disables 1M context at the runtime level. There is no
automatic fallback and no change to entitlement policy: extended context depends on
model and account availability, which the shipped docs link to.

On cost: Anthropic documents standard model pricing for the 1M window, so this is not a
move into a premium tier. Longer reviews can consume more total tokens; the existing
maxBudgetUsd: 25 per-consultation cap is unchanged.

The suffix does not reach the provider. Internal normalization (X5) re-attaches the
marker for concrete ids, so it is visible right up to the request — but the strip happens at
the API boundary itself, on both runtimes:

• 0.2.105 (cli.js): UT(q) = q.replace(/\[(1|2)m\]/gi, ""), applied at
  beta.messages.create({...P, model: UT(P.model)}), at countTokens, and on the bedrock path.
• 0.2.141 (native binary): the same function, minified as
  KL(H) = H.replace(/\[(1|2)m\]/gi, ""), applied as model: KL(A.model).

The context budget and the context-1m-2025-08-07 beta header are separate effects of the
suffix; the native binary carries that header (4 occurrences) and
CLAUDE_CODE_DISABLE_1M_CONTEXT alongside the same
rG(H) = /\[1m\]/i.test(H) gate.

8 files, +68/−25 (excluding the builder thread log).

Regression evidence

The tests were applied alone first, against the unfixed production code:

	Result
Test files only (no production fix)	13 failed / 345 passed
Full fix applied	358 passed / 358

Failures on baseline were exactly the expected ones —
expected 'claude-opus-5' to be 'claude-opus-5[1m]' and
Invalid model id "claude-opus-5[1m]" thrown from validateModelId.

Coverage: the shipped default, the SDK boundary (what actually reaches options.model),
configured ids including [1m], --model-id override resolution, the 200-character
total-length boundary (196+[1m] accepted, 197+[1m] rejected), and malformed suffixes
([1m] alone, opus[1m]extra, opus[1m][1m], opus[2m], opus[).

Verification

• pnpm build — passed.
• Full @cluesmith/codev suite — 5754 passed, 48 skipped, 0 failed (288 files).
• All 7 GitHub checks pass.
• git diff --check clean; the two consult.md copies verified byte-identical.
• validateModelId is the only production consumer of MODEL_ID_RE.
• Live confirmation (from the prior investigation, before this branch): with the [1m]
  default installed, the original PR afx send: add --interrupt-after <seconds> (bounded patience, then one forced delivery) #1640 review completed in 356.7s across 41 tool calls
  at 188,387 input tokens with no context error, and a separate bounded test processed
  465,431 input tokens in a single request.

🤖 Generated with Claude Code

[mohidmakhdoomi]

Architect integration review — preliminary (medium risk)

I read the full diff and verified that all eight seeded source/test/doc files match the handoff manifest. The change is appropriately scoped: select [1m], permit that suffix without weakening the total-length/character restrictions, preserve explicit overrides, and mirror the docs. No implementation blocker found so far. All seven GitHub checks currently pass; the required independent Claude integration review and final builder CMAP are still finishing.

Please correct the PR body and builder narrative, keeping the eight seeded files unchanged:

1. Do not describe all five original attempts as context failures or say Gemini reviewed successfully. One Claude attempt hit the separate usage limit; the original architect-side Gemini run wrote an unauthenticated skip notice, while Codex completed. This is already clarified in the issue discussion. Replace “any large review” with the observed repeated failures on PR afx send: add --interrupt-after <seconds> (bounded patience, then one forced delivery) #1640.
2. Scope the runtime claim to bare claude-opus-5 with SDK 0.2.105/0.2.141: that ID budgets 200K and its [1m] variant budgets 1M. The abbreviated window function has other branches for recognized models/settings, so “1M only with a suffix, 200K otherwise” is not a universal statement.
3. Describe the traced symptom as accumulated review-context exhaustion followed by Prompt is too long; avoid asserting an exact client-versus-server rejection mechanism or that all changed files were read in full. The saved sessions establish substantial prior tool activity, not that stronger mechanism.

These are factual corrections to the review record, not a request to expand the fix. This comment is not human PR-gate or merge approval.

[mohidmakhdoomi]

Architect integration review — Claude findings adjudicated

Risk: medium (10 files, shared model-ID validation; no protocol/state-machine implementation changes). The required independent Claude integration review completed in 254.0s with COMMENT, confirming the fix, validation boundaries, backwards compatibility, and contained scope. I checked its findings against the actual SDK rather than accepting them uncritically.

Disposition

1. Claim that [1m] reaches the provider as part of the model ID: rejected. The reviewer stopped at internal normalization (X5). In the locked SDK 0.2.105, I independently found:

   function UT(q) { return q.replace(/\[(1|2)m\]/gi, "") }
   // API request construction:
   { model: UT(A.model), messages: ..., ... }
   // Non-streaming API call:
   j.beta.messages.create({ ...P, model: UT(P.model) }, ...)

   Internal normalization retains the marker; API request construction strips it. The context budget and beta header are separate effects. The PR's suffix-stripping statement is correct. Anthropic documents the same provider-boundary behavior.
2. Account/provider availability: advisory, not demonstrated breakage. Live validation covers this environment, not every account/provider. The shipped docs already point to model/account availability, describe unchanged explicit overrides, and document CLAUDE_CODE_DISABLE_1M_CONTEXT=1. For clarity in this PR's record, please add the explicit opt-out: consult.models.claude: "claude-opus-5" retains the previous selection and 200K budget on the inspected runtimes. No automatic fallback or entitlement-policy changes are requested.
3. Version wording/reflow: non-blocking. The corrected PR body now names the inspected SDK versions; the source comment and mirrored docs accurately explain the older-runtime behavior. No need to rewrite the seeded patch for cosmetic reflow.

The earlier narrative corrections are appropriate. Do not amend or force-push old commits just to revise their historical wording; the corrected PR body and committed narrative are the durable correction. Keep all eight seeded files unchanged.

No implementation blocker remains from architect integration review. Final builder Claude CMAP and the narrative commit are still pending. This is a technical review disposition, not human PR-gate or merge approval.

[mohidmakhdoomi]

Final CMAP disposition — no implementation blockers

I read the final builder Claude review: APPROVE / HIGH, joining Gemini and Codex (APPROVE / HIGH). All eight seeded files still hash-match. No further source, test, or shipped-doc changes are requested.

A few advisory assertions in that review are incorrect; please correct only the PR/thread record rather than importing those assertions into the shipped docs:

• Disable-context behavior: UT(q) strips the marker unconditionally at request construction; it does not consult ZG() or the disable-context environment variable. Thus the claim that a concrete ID keeps [1m] on the wire when disabled is contradicted by the same API-boundary evidence already verified. The existing disable-context doc line needs no fix.
• Pricing: remove the new assertion that this necessarily triggers premium pricing above 200K. Anthropic's current documentation describes standard model pricing for the 1M window. Longer reviews can consume more total tokens; the existing $25 consultation cap is unchanged. That is the accurate, sufficient cost note.
• Stats: MetricsDB.summary() in metrics.ts groups by model (the lane), not model_id. Recorded concrete IDs change as intended, but the claimed split of consult stats lane totals does not occur.
• Runtime packaging: 0.2.141 uses a separately packaged native runtime, but its package.json pins every @anthropic-ai/claude-agent-sdk-<platform> optional dependency to exactly 0.2.141, and SDK resolution locates that package's binary. Do not imply this normally selects an arbitrary, unversioned system CLI. The 0.2.105 JS / 0.2.141 native-layout distinction is real; the latter is still SDK-distributed/versioned.

Keep this correction concise, commit the narrative once, and retain the already corrected PR summary rather than replacing it with the reviewer's generated summary (which repeats the original overclaims). These factual dispositions do not warrant another model review of the unchanged fix.

CI was all green on e339b399f; the review-log-only successor is rerunning CI. Human PR-gate and merge approval are still pending.

[mohidmakhdoomi]

Final CI — frozen head 23de98895

All 7 checks pass. PR state OPEN, MERGEABLE.

Check	Result	Duration
Artifact-Canvas Browser Tests	pass	48s
CLI Integration Tests	pass	1m7s
CLI Tests (macos-latest)	pass	58s
CLI Tests (ubuntu-latest)	pass	1m5s
Package Install Verification	pass	2m24s
Tower Integration Tests	pass	1m45s
Unit Tests	pass	2m45s

Review status

CMAP complete — all three lanes APPROVE / HIGH: gemini (49.1s), codex (148.0s),
claude (750.4s), no blocking issues from any lane. Architect integration review returned
COMMENT with no implementation blocker, and three of the final review's advisory
assertions were withdrawn after verification against the source (disable-context,
premium pricing, consult stats split).

State

• All 8 seeded files unchanged — 8/8 hash-match the handoff manifest.
• The three commits after the fix are narrative-only; no source, test, or shipped-doc
  change since b73c5d21d.
• Regression evidence unchanged: 13 tests fail with the test files applied alone against
  unfixed production code, 358/358 pass with the fix. Full suite 5754 passed / 48 skipped
  / 0 failed.

Head is frozen for human review. No gate approval or merge has been performed — awaiting
an explicit human decision.

[mohidmakhdoomi]

Final architect integration review — ready for human decision

Reviewed head 23de988. Medium risk: 10 files, +331/-25 including the protocol record and narrative; the functional patch remains the original eight files (+68/-25), all SHA-256-identical to the handoff.

• Problem: large Claude consultations repeatedly exhausted review context and produced no review artifact.
• Root cause: independently verified SDK/runtime 0.2.105 and 0.2.141 budget bare claude-opus-5 at 200K; [1m] requests 1M. Codev's validator rejected the suffix, blocking that override.
• Fix: explicit 1M default, narrowly extended model-ID syntax retaining the total 200-character bound, unchanged override precedence, regression tests, and mirrored docs. No unrelated protocol/retry redesign.
• Testing: 13 regression failures on baseline; 358/358 targeted tests green with the fix; full package suite 5754 passed / 48 skipped / 0 failed; build passed; diff whitespace check clean. All seven CI checks pass on this exact head. The installed implementation completed the original review and a separate live request with 465,431 input tokens.

Gemini, Codex, and Claude final CMAP verdicts are APPROVE / HIGH. The additional required Claude integration review returned COMMENT; its findings and the CMAP advisories were independently adjudicated in the preceding comments. Factual corrections are recorded, with no remaining implementation blocker.

Technical recommendation: merge after explicit human approval. This comment does not approve the human Porch gate. Builder retains ownership of gate execution, merge, and protocol verification; architect will close the issue after merge.

[mohidmakhdoomi]

The human approved the submitted fix in the architect conversation. Porch recorded that local PR-gate decision in 1fe1037. I independently verified its delta from the reviewed head is only approval bookkeeping in status.yaml; all eight seeded source/test/doc files remain unchanged. All seven CI checks pass on this final head.

We are contributors, not upstream maintainers. Our local sign-off is not GitHub maintainer approval. This PR remains open awaiting an eligible maintainer/reviewer approval and an authorized upstream merge. No self-review or branch-protection bypass is authorized.

The branch is frozen and the builder worktree is preserved. No extra narrative commit is needed; this comment records the disposition. Protocol completion, issue closure, and cleanup wait for a real upstream merge.