Add docs/Topics/Claude-in-CI.md, extracted from someones.computer #20

Merged
geusebio merged 1 commit from docs/claude-in-ci-principles into main 2026-08-13 12:43:14 +00:00
Owner

Adds docs/Topics/Claude-in-CI.md — 38 genericized principles for running Claude
headlessly inside a CI pipeline, extracted from someones.computer's
feature/ci-intake-read-commands branch: four Forgejo Actions pipelines (triage,
propose, review, resolve-conflicts) built up over ~60 commits of hard-won fixes.

Organized into six categories:

  • Safety and guardrails (kill switches, credential scoping by phase, deny-list granularity)
  • Cost and rate control (fail-closed budget gates, per-invocation spend ceilings, durable ledgers)
  • Reliability (resource-kill sampling, model-fallback error taxonomy, output durability)
  • Evidence and auditability (tri-state verification claims, fail-closed trust gates, transcript tiers)
  • Testing the harness itself (stub/fixture testing, local CI runs, harness bug libraries)
  • Other (binary pinning, credential handling, AI-review-of-AI-work bias)

Stripped of the source system's specifics — PHP/Symfony tooling, its exact tuned
numbers, its runner image identity, its issue-tracker taxonomy — in favor of the
underlying mechanism. A closing "What not to port" note says so explicitly.

Optional, adopted the same way as Documentation-Conventions.md: relevant only to a
project building or running this kind of unattended-agent-in-CI automation; interactive
Claude Code use needs none of it.

No code, no tests — one new doc plus one README table row.

Adds `docs/Topics/Claude-in-CI.md` — 38 genericized principles for running Claude headlessly inside a CI pipeline, extracted from `someones.computer`'s `feature/ci-intake-read-commands` branch: four Forgejo Actions pipelines (triage, propose, review, resolve-conflicts) built up over ~60 commits of hard-won fixes. Organized into six categories: - Safety and guardrails (kill switches, credential scoping by phase, deny-list granularity) - Cost and rate control (fail-closed budget gates, per-invocation spend ceilings, durable ledgers) - Reliability (resource-kill sampling, model-fallback error taxonomy, output durability) - Evidence and auditability (tri-state verification claims, fail-closed trust gates, transcript tiers) - Testing the harness itself (stub/fixture testing, local CI runs, harness bug libraries) - Other (binary pinning, credential handling, AI-review-of-AI-work bias) Stripped of the source system's specifics — PHP/Symfony tooling, its exact tuned numbers, its runner image identity, its issue-tracker taxonomy — in favor of the underlying mechanism. A closing "What not to port" note says so explicitly. Optional, adopted the same way as `Documentation-Conventions.md`: relevant only to a project building or running this kind of unattended-agent-in-CI automation; interactive Claude Code use needs none of it. No code, no tests — one new doc plus one README table row.
38 genericized principles for running Claude headlessly inside a CI
pipeline, pulled from a production system (four Forgejo Actions pipelines:
triage, propose, review, resolve-conflicts) after ~60 commits of hard-won
fixes. Organized into six categories: safety/guardrails, cost/rate control,
reliability, evidence & auditability, testing-the-harness, and other.

Stripped of the source system's specifics: PHP/Symfony tooling, its exact
tuned numbers (run caps, spend ceilings, parallelism limits), its runner
image identity, and its issue-tracker taxonomy. What's left is the
mechanism, not the calibration — a "what not to port" note says so
explicitly.

Optional, like Documentation-Conventions.md: only relevant to a project
building or running this kind of automation.
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
Grey.ooo/agent-standards!20
No description provided.