[lukemaj]

What changed: a ContextMD deny that Grok showed cut (the rule's start, then … [+N chars]) is now logged truncated: it counts toward the deny cap, and an inline rule stays armed because it was not seen whole.
Why: Grok shows a hook's deny reason only up to 256 characters; the evidence check required the whole rule in the projected output, so those denies were logged hidden, never counted toward the cap and kept denying (chromeria#228; live sample: thread mcp:843cf768, 256 characters shown, then … [+54 chars]).
So what: the planner runs the Sol and Grok reviews; chromeria#229 then makes Grok denies fit, so the target path is never the part cut.

Closes #228

Change

ServingState.applyChecks: when the whole probe is not in the projected output, shownCut checks for Grok's cut mark and that the text before it holds the probe's opening (its first 40 characters) followed only by more of the probe. Such a deny is adjudicated shown with whole = 0, outcome truncated, and denyShown counts it; the inline lift needs whole = 1, so it does not lift. Anything else stays hidden as before.

Upstream: fork-only module, no upstream counterpart. Independent of #226 (different function); whichever merges second rebases.

Proof (head efb064cc89)

• New test "counts a deny Grok showed cut as shown and truncated, not hidden (contextmd serving: a shown Grok deny is logged as hidden, so the deny cap never counts it #228)": failed before with hidden; now truncated, the inline rule stays denied, and the cap releases the fourth call.
• vp test run src/contextmd/: 68 passed; lint, format and apps/server typecheck clean.
• contextmd contract suite against this branch: 17 OK.

Elon record

• Requirements and who asked: the log says honestly whether the model saw a rule, and the cap counts shown denies (contextmd#3, user decision 2026-10-12); filed as chromeria#228.
• Deleted: nothing.
• Bottleneck: the evidence comparison.
• Checked myself: the live Grok projection item (256 characters, then … [+54 chars]) and the tests above.

Work by Claude Opus 5.5 in Claude Code (Chromeria thread).

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.9 KiB	20.9 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	−7 B (−0.1%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	−7 B (−0.6%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.2 KiB	21.2 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: 58abe11 · PR result: efb064c · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[lukemaj]

What: Independent review PASS at efb064c.
Why: No blocking findings in this slice.
So what: The planner can use this exact-head review with required companion review and delivery dependencies.

VERDICT: PASS
Base: 58abe11.

No blocking findings. A projected Grok deny prefix followed by the cut marker is adjudicated shown with whole=0, logged truncated, and counted toward the cap without lifting its inline rule. The projection consumer recursively extracts text from finished structured dynamic-tool outputs, excluding input.

Proof:

• Scoped serving suite: 68/68 pass; unchanged head tests on merge-base production sources: 1 assertion failure, 67 pass. The new case is hidden before, truncated after; the inline rule remains armed and the fourth call releases.
• ContextMD Observer MCP toolkit: agents query task cost and thread diagnostics without a Skill #50 full Python suite against this head: 114 pass, including 17 real-server/hook contracts.
• Hidden-other-hook, shared-prefix whole-evidence, pending-evidence and read-lift controls pass.
• Exact-head CI is green, including all server shards and Check.

Full-read tightening is a separate slice in #226, not included or regressed here.

Reviewed the complete diff, linked Issue and prior PR/Issue review comments, and affected selection, harness, state and transport consumers. All eight heads matched the brief. Tests used exact-head git archives in temporary directories and existing dependencies. Baseline runs held the head tests constant and restored changed production files from the stated merge base.

Chromeria command: vp test run --config review.config.ts, with include=['src/contextmd/*.test.ts'], node environment and serial files.
ContextMD command: PYTHONDONTWRITEBYTECODE=1 CONTEXTMD_CHROMERIA= python3 -m unittest discover -s tests -t .; baseline: python3 -m unittest tests.test_serve.

Passing state/route tests retain the cap of 3, failed/pending reads, session/epoch isolation, dropped/concurrent deliveries, fail-open guards and after-action suppression, including Grok carries. No checkout, reset, pull or source edits in any checkout. No private data or credentials found in the diff. Proof is synthetic hook/production-code execution, not a fresh live model session or deployment.

Reviewer: Codex gpt-6.1-sol, independent review, requested high effort.

[lukemaj]

What: PASS at efb064cc89f241c8f7b577771eab479bda4248e5. No blocking findings.
Why: A Grok deny whose projection ends in the cut mark is recorded as shown and truncated, so it spends the cap and does not lift the rule.
So what: This candidate can proceed on its own. It does not change deny wording; chromeria#238 is what keeps the target inside the 256 bytes Grok shows.

Head matches the brief. Base is main.

shownCut in ServingState.ts requires the unicode ellipsis and [+N chars], then accepts the text only when the bytes before that mark are a prefix of the delivered probe. A whole probe still takes the existing includes(probe) path. Hidden stays the path for a deny the model did not see, and that path still does not spend the cap.

The mark matches a live Grok deny observed this session (… [+N chars]). DENY_CAP remains 3. A cut inline rule stays armed until the third shown deny. Fail-open and after-action suppression are unchanged.

CI on this head: Lint, Test, Test Server 1-6, and Check succeeded. I did not re-run the server suite locally.

No private data in the diff.

[lukemaj]

Merged by the planner: review/independent-sol and review/independent-grok both success on efb064c, checks green.