feat(sidecar,shell): 24-step turns, failure output reaches the model (v0.5.1 HGW-W5/W6)

[SaulBuilds]

HGW Week 1, runtime, 4 of 5. Stack: #79 → #80 → #81 → this PR (base hgw/w1-project-create) → #83. Green on its own. Refs #76 (HGW-W6) and CitrateNetwork/citrate-core#291 (HGW-W5, runtime default).

What changes

• Turns: the default maxSteps goes from 8 to 24 (sessions::DEFAULT_MAX_STEPS); the cap stays 32. A session's own maxSteps still wins. The sidecar-wide default is CITRATE_HERMES_MAX_STEPS (clamped 1–32).
• Capture (agent-shell): a capped stream now keeps its first half and its last half (the first error and the final cause), with the truncation marker between them. Before, it kept only the head.
• shell_run on failure: a non-zero exit puts the exit status and the end of stderr in the summary the model reads first. If stderr is empty, it uses the end of stdout. The tail is at most FAILURE_TAIL_BYTES = 2048 bytes. The run facts carry stderr_tail / stdout_tail. Successful runs are unchanged.
• Toolchain with no report (for example forge exiting before tests):
  • the summary now states the exit code;
  • recognized compiler errors stay sanitized and trusted, as before: injected source text never reaches trusted output;
  • when no cause is recognized (a missing solc or library, a crash), the end of stderr is handed over raw and bounded, and the result is Untrusted. It taints the session; with feat(sidecar): taint escalates to HIC approval (v0.5.1 HGW-W1) #79, later actions then go to cards.

Interface for the core lane

• core#291: build sessions can stop sending maxSteps, or send 24; 32 is the most.
• shell_run results carry run.stderr_tail / run.stdout_tail on failure. The Activity card can show them.

Not done from #76

Trimming the model's copy of shell_run output to 4 KB while the event log keeps the full text. Model content and event content are one string today, so this needs a loop change. The summary tail gives the model the cause either way.

Tests (red first)

• agent-shell runner: +1 (head and tail kept); two capture assertions extended.
• loop_capability_tests: +3 (24-step default, per-session and per-sidecar settings, environment parsing).
• shell_run failure tail: +1.
• Toolchain unparsed stderr: +1 (Untrusted with the raw tail; recognized diagnostics stay trusted).
• Workspace: 2282 → 2288 passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_018wWZZ8GRVU9USh3kKQHsFe

[BerryManifold]

Reviewed step-limit defaults/clamping, bounded head/tail output capture, UTF-8 boundaries, and failure-output trust handling. Independently ran all 3 loop-capability tests and 3 capped-output runner tests on the #84 stack head: passed. No new blocking finding in this incremental diff. Approval does not clear the separate git-state blocker affecting the stack or replace required CI checks.

[Matr0xshka]

Reviewed step-limit defaults/clamping, bounded head/tail output capture, UTF-8 boundaries, and failure-output trust handling. Independently ran all 3 loop-capability tests and 3 capped-output runner tests on the #84 stack head: passed. No new blocking finding in this incremental diff. Approval does not clear the separate git-state blocker affecting the stack or replace required CI checks.

[SaulBuilds]

Landed in main via #84 (squash 75e21cd), per the owner's 2026-10-15 decision to land the HGW stack #80–#84 as one squash. Reviewed and approved here.