fix(ai): honor the documented GJC_NO_STRICT bypass

[10kH]

Another documented-but-unread name, found by checking the canonical env reference against what other docs teach.

Problem

docs/ai-schema-normalize.md:32
  adaptSchemaForStrict ... consults `GJC_NO_STRICT` (env `GJC_NO_STRICT`) for the global bypass.

packages/ai/src/utils/schema/adapt.ts:14
  export const NO_STRICT = $flag("PI_NO_STRICT");   // documented name never read

Same PI_ → GJC_ drift class as #2827 / #2943 / #3041 / #3171 / #3245 / #3296, and like #3296 the drift costs something concrete: this flag is the escape hatch that turns off OpenAI-style strict function-schema enforcement. An operator working around a provider that rejects strict schemas sets the documented name, and strict mode stays on.

Fix

$pickflag("GJC_NO_STRICT", "PI_NO_STRICT"). $pickflag applies the same boolean contract $flag already used — first non-empty key decides, case-insensitive 1/true/yes/on — so legacy behavior is unchanged and the canonical name simply takes precedence.

The boolean contract is right here, unlike #3296: this is a bypass, so GJC_NO_STRICT=0 meaning “stay strict” is the safe direction. In #3296 the same shape would have let =0 silently re-enable reading a token out of another application, which is why that one stayed presence-based. Calling that out because the two look superficially identical.

Also: a reference-page gap

GJC_NO_STRICT was taught only in the schema-normalisation note. docs/environment-variables.md opens by saying it is “derived from current code paths” and “documents only active behavior”, so a knob read in packages/ai/src/** belongs there. Added.

Tests

no-strict-env.test.ts runs each scenario in its own process, because NO_STRICT is a module-level constant: strict on by default, each name honoured, the documented boolean spellings accepted case-insensitively, an explicit falsey value respected, and the canonical name winning over a truthy legacy value.

Proof-first: reverting to the legacy-only read fails three of the six.

Verification

bun run check in packages/ai (whole-package biome + tsc): exit 0. New suite: 6 pass / 0 fail. Strict-named suites: 277 pass / 0 fail. Docs index regenerated; check:public-sync in sync.

[Yeachan-Heo]

Verdict: APPROVE — exact head 7709d367764b8b21a4a793ed822b911ee0a52065

Same PI_ → GJC_ documented-but-unread drift class as #2827/#2943/#3041/#3171/#3245/#3296, gating strict-schema enforcement bypass. Documented GJC_NO_STRICT alias verified as a genuine, correctly-behaving boolean.

Checks performed:

• Pre-fix defect confirmed: docs/ai-schema-normalize.md:32 documents GJC_NO_STRICT; pre-fix adapt.ts:14 read only $flag("PI_NO_STRICT") — the documented name was a silent no-op, matching the PR's claim.
• $pickflag semantics verified directly (packages/utils/src/env.ts:284-290): iterates keys in order, returns the boolean value of the first key with a non-empty value (not first-truthy) — i.e., an explicit GJC_NO_STRICT=0 correctly wins over a set PI_NO_STRICT=1 and resolves to false, matching the PR's "canonical name wins over a truthy legacy value" test case exactly.
• Boolean-vs-presence contract choice justified: unlike #3296 (GJC_AUTH_NO_BORROW, presence-based because =0 re-enabling token borrow would be the unsafe direction), here GJC_NO_STRICT=0 meaning "stay strict" is the safe failure direction, so routing through the existing $pickflag boolean contract (rather than a bespoke presence check) is the correct choice — the PR explicitly calls out and reasons through this distinction rather than copy-pasting the #3296 pattern blind.
• Docs reference-page gap: GJC_NO_STRICT added to docs/environment-variables.md's OpenAI-code table, consistent with that file's stated scope ("derived from current code paths… documents only active behavior").
• Tests: hermetic — each scenario spawns a fresh process via no-strict-probe.ts since NO_STRICT is a module-level constant, correctly avoiding a stale-import false pass. Six cases: default-strict, documented-name honored, legacy-name honored, case-insensitive boolean spellings, explicit falsey documented value, and canonical-wins-over-truthy-legacy. Proof-first claim (reverting to legacy-only fails 3/6) is consistent with the diff.
• CI: all applicable checks pass on the exact head, including the new test:packages/ai/test/no-strict-env.test.ts job explicitly green.
• Novelty: no other open PR touches adapt.ts's NO_STRICT resolution; not a duplicate.
• No automated reviewer comments present.

No merge action taken — read-only review verdict only.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Verdict: APPROVE — exact head d0efab7e95bdd58e254423a21ff8b8bf97ce780e

Rebase-only refresh of the prior exact-head review at 7709d367764b8b21a4a793ed822b911ee0a52065 (APPROVE).

Checks performed:

• Content identity confirmed, not assumed: diff <(git show 7709d367 --format= -- docs/environment-variables.md packages/ai/src/utils/schema/adapt.ts packages/ai/test/fixtures/no-strict-probe.ts packages/ai/test/no-strict-env.test.ts) <(git show d0efab7e --format= -- docs/environment-variables.md packages/ai/src/utils/schema/adapt.ts packages/ai/test/fixtures/no-strict-probe.ts packages/ai/test/no-strict-env.test.ts) is empty — the PR's own commit is byte-for-byte identical to the version the prior review validated. CHANGELOG.md is intentionally excluded from this file set: its diff context shifted because independently-merged dev commits inserted new entries above the same insertion point, which is expected rebase churn, not a content change to this PR's own line.
• Rebase mechanics verified: git merge-base confirms this PR carries the same single commit both before and after the rebase; every additional commit between the two heads is an independently-merged dev commit (own-PR-numbered), not new work folded into this PR.
• Exact-head terminal CI: 17 success / 0 failure / 0 pending / 5 skipped (total 22). No pending/queued checks remain.

Prior technical analysis still applies verbatim to this head.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

MERGE_READY (verdict reconciliation — exact head)

Reattaching the existing durable verdict to the current exact head 824439238ddd2211f023b2c24ffc7bc1d8503216 (base dev @ fbb5670709a1b1e04abddd1cc4827143fb63b876). This is a reconciliation, not a new review: the prior APPROVED was submitted at d0efab7e, and the head moved by rebase only.

Evidence:

• Normalized patch of d0efab7e and 82443923 (file list + hunk bodies, hunk headers excluded) is byte-identical — SHA256 5a95ab30… for both. File-level shape is unchanged: docs/environment-variables.md +1-0, packages/ai/CHANGELOG.md +1-0, packages/ai/src/utils/schema/adapt.ts +2-2, packages/ai/test/fixtures/no-strict-probe.ts +5-0, packages/ai/test/no-strict-env.test.ts +59-0, packages/coding-agent/src/internal-urls/docs-index.generated.ts +1-1.
• Only the parent changed (b853f15d → fbb56707). No source, test, or docs delta.
• Exact-head checks: 17 success, 6 skipped, 0 failure. mergeable_state: clean.

Verdict carries forward unchanged. Merge remains gated on dev returning green (current dev @ fbb56707 is red).

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Verdict: APPROVE — exact head 49be461ef0ba39486fded9987d206dcfa297db56.

Behavior and release-surface revalidation:

• NO_STRICT now resolves with $pickflag("GJC_NO_STRICT", "PI_NO_STRICT"), so the documented canonical name works, the legacy alias remains supported, and the first non-empty canonical value wins even when explicitly falsey. That safe-direction boolean contract is appropriate for a bypass flag.
• Six fresh-process scenarios cover default strict mode, both names, case-insensitive truthy spellings, explicit false, and canonical-over-legacy precedence.
• The semantic rebase lands on current dev parent f3a34362243e9ddde049844094e52547557328f9. The changelog entry is under Unreleased; the environment reference was merged with current content and the internal docs index regenerated from source rather than conflict-resolved by hand.
• Local focused suite: 6/6 passed. AI typecheck, Biome, generated-doc public-sync, and diff checks passed.
• Exact-head CI is terminal with every reported check SUCCESS or SKIPPED, including the focused NO_STRICT test plus both AI and coding-agent checks.
• Mergeability is MERGEABLE / CLEAN, and the exact remote head/base were re-pinned immediately before this verdict.

— GJC provider/trust review, 2026-07-28