gstack/docs/evals/headroom-diagnosis-2026-10.md

Headroom diagnosis, October 2026 (A1-A3)

Per-phase timings for the cases the reliability plan named as slow, measured from stored CI artifacts of census 37198445662 (head fd6854b) and, for the one red, 37176837432. Native captures carry message timestamps; PTY cases carry the assistant-message timestamps their observation files keep. Durations in the trial history (wave-trials) for multi-test files are the JUnit sum of tests that run concurrently, not wall time or session time, so they overstate headroom for plan-mode-no-op and plan-ceo-mode-routing.

A1. ship-docsync-completion and -store (session budget 585 s)

History: completion 349-517 s, store 377-466 s across the eight wave censuses.

Phase (37198445662)	completion	store
Prepare: read phase, skills, Git state, fixture files; snapshot	85 s (two thinking gaps of 9 s and 26 s)	99 s (33 s thinking)
Compose child prompt (9.5 KB / 7.1 KB Agent input)	43 s	32 s
Child (document-release, foreground)	122 s	118 s
Validate (compare, Git reads, small file reads)	3 s tool time, 38 s thinking	3 s tool time, 13 s thinking
Write records after the child	45 s (section copy, report, result JSON with typed hashes)	78 s (child output copy, section, record, report)
Publish stand-in, verify, final message	32 s	25 s

Tool time in the parent is about 1 s per session. Reading release-file contents does not dominate (10 small files, ~1 s), so the parent keeps its reads and the child result contract stays unchanged. The parent's own model time dominates (~270 s of ~390 s). Two parts are cuttable without weakening the gate:

• Records after the child (product): the parent retyped hash maps and child output into several files. The gate now records post-child hashes by rerunning the gstack-docs-candidate snapshot command (--out <audit-id>-post.json), Step 16 runs compare on it, and each record is saved once and cited by path.
• Child prompt copy (fixture only): ~70% of the child prompt is the fixture's observation interface, which the fixture tells the parent to copy into child prompts (docsNativeInterface, "include this interface in child prompts"). Writing that interface to a fixture file the child reads would save ~30-40 s per session; it touches the shared detector/actor module, so it is a lane D request.

A2. plan-mode-no-op and plan-ceo-mode-routing

plan-mode-no-op is not a headroom problem. Its six PTY sessions start concurrently; 37198445662's file wall is 263 s (slice entry 271.6 s) while the JUnit sum is 518 s. The slowest session is 159 s of its 300 s observation budget (53%). It runs in gate slice 4, not the 711 s critical-path slice 11.

plan-ceo-mode-routing runs two concurrent sessions (wall 378 s, sum 687 s). The slowest session (HOLD SCOPE) is 378 s of 600 s (63%). The tight budget is the test's 240 s post-selection window:

HOLD SCOPE session (37198445662)	elapsed
Preamble	17 s
Pre-review audit batch	46 s
Setup questions and answers	45 s
Learnings search, Step 0 analysis written to the plan	50 s
Mode question rendered and answered	20 s
Handoff helper, plan write, two section reads, Codex probe, then posture chat	114 s
HOLD continuation question and posture assessment	76 s

In the red (37176837432) the session skipped the handoff helper and paraphrased the mode ("holding scope"), then spent ~60 s per decision on the ledger save/read-back/ask cycle until the 240 s window closed. The skill now sends the handoff line before any other tool call, plan write or section load, and never skips the helper. The remaining time is model latency of the decision procedure (about 60 s per question: comparison grid, save, Read-back, ask); cutting that would weaken the question-persistence gate, so it stays a named item.

A3. Census critical path

Gate slice 11 (711 s, the census critical path) holds 17 short files. Their recorded estimates (scripts/paid-test-durations.json, recorded 2026-09-29) total 1078 s; they ran 1353 s. Four files ran about twice their estimate in every wave census:

File	estimate	wave census range
skill-e2e-plan.test.ts#plan-review-report	76 s	168-273 s
skill-e2e-plan-ceo-plan-mode.test.ts	37 s	85-120 s
skill-e2e-deploy.test.ts#benchmark-workflow	62 s	59-121 s
skill-e2e-safety-codex-consult.test.ts	83 s	111-212 s

None is near its own budget (plan-review-report peaks at 46% of 600 s). The fix is a refreshed duration seed so the planner packs slice 11 within the 540 s slice budget (lane F / integrator). review-army-perf and plan-design-review-plan-mode need no work.

plan-eng-multi-finding-batching: 975 s in census 37228573062 (GSTA-23 branch)

The red trial ran 975 s of its 1,500 s armed session (65%), against 313-409 s in the nine earlier censuses (37162480720 … 37198445662). The wall is a consequence of the detector miss, not slower work: the runner stops collecting once the counter credits FLOOR = 3 review questions (normally by D3, about 5-7 minutes). The counter credited none, because the report declared its target mid-line ("… by the user. Review target (fixed): PLAN.md …"), so collection ran on through D10 and the completed report. With the target-field fix the same capture is credited at D5 (521 s). Replay: bun test test/detector-corpus-plan-eng-multi-finding-batching.test.ts. No budget or kind change.

review-army-simplification: 159 s of 180 s in the Oct 12 census (#3153)

The pass-rate gate pooled 159 s of the 180 s session budget (88.1%, over the 85% cap); census 38212265208's own panel ran 142, 91 and 146 s. Phases from the three transcripts (paid slices 25-27):

Phase (38212265208)	t1	t2	t3
Read the extracted review-SKILL.md, checklist, specialist; search for a "## Step 4.5" heading the carved fixture no longer has	17 s	15 s	36 s
Scope and stats commands, analysis, and reconciling the prompt's "over 100 lines" with the 68-line diff	64 s	29 s	13 s, then a subagent (29 s) and re-verification (21 s)
Compose the unbounded report	45 s	33 s	31 s
Final message (each restates the findings and the premise mismatch)	11 s	8 s	11 s

The fixture, not the skill, carried the extra work. The case now uses the staging its siblings use (stageReviewArmySession: recorded scope, stats, learnings and diff; only review-army.md and simplification.md staged), selects the specialist with the real /review --simplification force flag, applies the checklist in-session and bounds the report to 30 lines. Budget and assertions are unchanged. Free check: the simplification scenarios in test/review-army-budget.test.ts.

ship-exploratory-late-input: a probe batched with its Reads (#3153)

Not a headroom case; recorded here because the cause is also fixture work. Census 38212265208 failed it with "probe preceded resource read". Capture 001 went out in the same provider response as the Reads of the caller's Step 8 and Step 9 sections, before any QA section was read (the Oct 7 capture mtb1007a batched it with the exploratory Read). /ship's prose never names the QA capture helper; the fixture prompt supplies it, and the fixture's scheduling contract (added 2026-10-06) invited batching "Reads and read-only discovery" without listing Read results among the dependencies a probe must wait for. The case passed on Oct 3 and Oct 5, before that contract. The contract now classifies captures, guard starts and runs as probes, never discovery; the ordering rule itself stays in /ship's Step 9 STOP, which the case still has to follow unaided. Free check: the census 38212265208 replay in test/qa-exploratory-callers.test.ts.