Skip to content

feat: implement deepcycle deep-learning skill - #3

Closed
aborzin wants to merge 1 commit into
danshapiro:mainfrom
aborzin:main
Closed

feat: implement deepcycle deep-learning skill#3
aborzin wants to merge 1 commit into
danshapiro:mainfrom
aborzin:main

Conversation

@aborzin

Copy link
Copy Markdown

No description provided.

@aborzinaborzin closed this Mar 17, 2026
@aborzinaborzin reopened this Mar 17, 2026
@aborzinaborzin closed this Mar 17, 2026
christianbauer-deltaorbit pushed a commit to christianbauer-deltaorbit/trycycle that referenced this pull request Apr 26, 2026
Sets explicit `# trycycle-step: timeout-seconds: N` headers on the 10
subagent templates listed in the spec. Together with the new lifesigns
subprocess probe (de68540) and the streaming runner (b2fa00b), the
default-everywhere posture is replaced with empirically-justified
per-template caps. Long healthy steps now have the headroom they
legitimately need, and short steps fail fast when they over-explore.
Values (matching the spec's recommendations and rationale):
template seconds category
----------------------------------------------------------------
prompt-executing-finalize.md 14400 long-iter
prompt-executing-next-task.md 10800 long-iter
prompt-test-strategy.md 3600 reasoning
prompt-test-plan.md 3600 reasoning
prompt-post-impl-review.md 3600 reasoning
prompt-planning-initial-scaffold.md 1800 creative
prompt-planning-initial-detail.md 1800 creative
prompt-planning-initial-commit.md 1800 creative
prompt-planning-edit.md 1800 review
prompt-planning-initial-survey.md 600 deterministic
Templates left without a header:
- prompt-executing-load.md (load is fast checklist work; not in spec
scope but a candidate for a follow-up tight-cap of ~600 s).
- prompt-executing.md and prompt-planning-initial.md (legacy single-
shot templates used via --single-shot; fall through to the runner's
EXECUTING_TIMEOUT_SECONDS / DEFAULT_TIMEOUT_SECONDS).
Tests:
- tests/test_step_timeout_header.py:
- replaces the narrower test_executing_next_task_template_declares_
long_timeout (which asserted both at 7200) with a comprehensive
test_shipped_templates_declare_expected_per_step_timeouts that
audits all 10 headered templates against an inline expected-value
table. Updating the map deliberately requires touching this test.
- adds test_header_wins_over_explicit_cli_timeout_seconds: dispatches
a fixture with header=1800, asserts header wins both with and
without --timeout-seconds 600 on the CLI; also asserts a
headerless template + --timeout-seconds 600 uses 600. The test's
failure-message names the resolver behaviour so a future flip is
a deliberate, documented change.
SKILL.md adds a "Per-step timeouts (template headers)" subsection
under "Sequenced phases" documenting the convention and the
how-to-choose-a-value note (deterministic vs creative vs reasoning vs
long-iterative).
Spec discrepancy worth flagging:
- The spec's acceptance criterion danshapiro#2 asserts "the CLI override takes
precedence over the header" and the spec's regression danshapiro#3 asserts
CLI=600 + header=1800 → 600 (CLI wins). That contradicts the
resolver shipped in commit 1987c89, where the HEADER wins and CLI
is fallback. Per the spec's primary directive ("Pure template edits
— no orchestrator code changes") I have NOT flipped the precedence.
The new test instead codifies the actual current behaviour with a
failure message that flags the discrepancy. If the spec's intended
precedence (CLI wins) is the true target, that's a 1-line resolver
change in _resolve_step_timeout_seconds plus updating two existing
tests' assertions — happy to do that as a follow-up.
Regression: 205 -> 206 tests passing; 5 skipped unchanged.
https://claude.ai/code/session_01XV4wwck5YZ9fxabteSeSk2
Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

@aborzin