feat(nano): round 3 — long-running work support + operational verification (FOR REVIEW)

[agentforce314]

Held for your review — will not be auto-merged. Reapplies the round-3 changes (#897/#898, reverted from main by #899 pending your decision).

Gap analysis driving the round (tb21-nano-flash-max-2, 25 incorrect tasks)

• 16/25 were wall-clock kills, mechanism measured: bash's 2-min default / 10-min cap forces long builds/training into nohup-and-poll loops — 225 poll-pattern commands across the failures; compile-compcert burned 111 polls, 307 bash calls, and died at the 300-turn adapter ceiling mid-build; caffe-cifar-10 (163 calls), train-fasttext, make-mips-interpreter share the signature. Each poll costs a full max-thinking turn; pi pays zero turns for a 25-minute make.
• 4 of 9 completed-wrong tasks were one-assertion near-misses on constraints stated in the task text (wrong source archive; primer Tm 6.5°C vs stated ≤5°C; clean-HTML entities mutated after claiming "fully verified"; untested CLI output format).

Changes

1. Bash: no timeout ceiling under nano — no default timeout (24h supervision deadline; abort + task wall-clock are the backstops), no 10-min cap on explicit timeouts, 1000ms floor kept, stock byte-identical. Doc + one anti-poll guideline ("ONE blocking command, tee to a log").
2. Runner turn ceiling 600 (NANO_MAX_TURNS override).
3. Operational verification guideline — extract stated requirements/constraints into a checklist, test the artifact against each.

Evidence

5 new tests; 70 nano + 125 bash regressions pass; live sleep 130 completes in one nano call (old default killed at 120s). Targeted k=1 validations on compile-compcert / train-fasttext / caffe-cifar-10 are running with this branch's wheel — results will be posted as a PR comment.

🤖 Generated with Claude Code

[github-actions]

Test Results

     5 files     997 suites   28m 54s ⏱️
15 321 tests 15 297 ✅ 19 💤 5 ❌
30 916 runs  30 853 ✅ 58 💤 5 ❌

For more details on these failures, see this check.

Results for commit 447446e.

♻️ This comment has been updated with latest results.

[agentforce314]

Targeted validation results (k=1, previously-failed tasks, this branch's wheel):

task	v2	v3 (this branch)	behavior change
compile-compcert	0.0 — 307 bash calls, 111 polls, died at 300-turn ceiling mid-build	1.0	4 polls, 283 turns, $0.075
caffe-cifar-10	0.0 — 163 calls, 16 polls, killed at 3601s	1.0	blocking training runs
train-fasttext	0.0 — 301-turn poll spiral (v1)	0.0 (still)	zero polls, 21 calls, proper blocking tee'd training — now compute-bound at the 3600s task budget: mechanism fixed, task is capability-bound

Also added since the PR opened: the run_in_background schema-trap removal (29 stranded launches in v2) and the stuck-command detection layer — concurrent pipe drain via raw os.read (fixes a latent all-modes deadlock: >64KB of output filled the un-drained pipe and blocked the child forever, which the stock hard timeout misreported as 'timed out') plus a nano-only output-idle watchdog (no output for 600s while running → killed with a teaching error; noisy long builds never trip it). 11 new tests; 201 nano+bash tests green.

Launching the full-suite tb21-nano-flash-max-3 with this branch's wheel.

[agentforce314]

Full-suite v3 result (tb21-nano-flash-max-3, this branch's wheel + vision + websearch):

Topline: 61/87 verified (+2 trials lost to a pre-agent infra failure — curl SSL error fetching the uv installer; zero agent involvement; being re-run) at $1.04. v2 was 64/89, v1 60/89.

Mechanism attribution is clean in both directions:

• All 6 gains are round-attributed: compile-compcert (validated fix, confirmed in-suite — 111 polls → 4), schemelike-metacircular-eval + path-tracing-reverse (razor-edge timeouts, saved poll turns), build-pov-ray + pytorch-model-cli + install-windows-3.11 (constraint-checklist class).
• Zero losses are round-caused: the idle watchdog fired 0 times all run; no over-commitment-to-blocking-call signature (7–69s/turn everywhere); the 9 losses = 2 infra deaths + 5 chronic k=1 flappers (tune-mjcf, regex-chess, feal…) + 2 fast-wrong answers.

Cross-run stability (the honest 0.8 frame): pass@1 = 60/64/61 across three draws (±3 noise); stable core v2∩v3 = 55; pass@2 = 70/89 (78.7%); pass@3 = 73/89 (82.0%). 16 tasks never solved in any run — video/vision-heavy, cross-toolchain builds, compute/think-bound — the genuine capability frontier at this model tier.

Verdict: the branch's mechanisms are individually validated and net-positive; single-draw pass@1 sits ~70±3% with the capability union above 82%.

[agentforce314]

Infra-rerun correction: both pre-agent infra deaths (curl SSL fetching the uv installer, zero agent involvement) pass when actually run — hf-model-inference 1.0, qemu-startup 1.0. Corrected v3: 63/89 (70.8%).

Final cross-run picture for this branch: pass@1 = 60 / 64 / 63 across the three draws (±2 around ~62); stable core 55; pass@2 = 70/89 (78.7%), pass@3 = 73/89 (82.0%); 16 never-solved tasks form the capability frontier at deepseek-v4-flash. All 6 v3 gains are branch-attributed, zero losses branch-caused, idle watchdog 0 misfires in 89 tasks.

[agentforce314]

Added: pi-style rolling-tail truncation + full-output spill file (447446e, requested follow-up)

The one residual bash-layer advantage pi held: its truncation keeps the tail of long output (where a build's error and final metrics live — stock head-keep drops exactly that) and spills the complete stream to a temp file whose path is handed to the model, so it can grep the whole log instead of re-running a 25-minute build.

Mechanics (port of pi's shell-output.ts):

• Reply = last BASH_MAX_OUTPUT_LENGTH chars + footer: [Showing last 698 of 4001 lines (165.8KB total). Full output: /tmp/bash-stdout-….log]
• Lazy-lossless spill: memory may hold up to 2× the limit before trimming; the file is created the moment the total crosses 1×, while memory still holds everything — complete from byte 0.
• Bounded memory: ~2× limit per stream (rolling deque) — a multi-GB log no longer accumulates unbounded in the drain's chunk lists.
• UTF-8 re-alignment at the trim cut (pi's trimToLastUtf8Bytes), stderr spools separately, all four result paths (normal/interrupted/timeout/idle-killed) render the footer.
• Nano-gated: stock keeps unbounded capture + head-keep, byte-identical (stock suites 142 passed).

Validation: nano suite 81 passed; live smoke with a 4001-line fake build log → reply ends with the FATAL line + footer, spill file complete (169,826 bytes). Nano Bash tool doc updated to state the tail+file contract (still <500 chars, pi-length guard green).