wt
b1014f36c9
Changes: Keep startup instructions and exact context packets inline while listing supplementary task inputs as mandatory project reads. Preserve root containment, missing-artifact guidance and budget failures. Add S21-006 exact UI context regression and cover claim/review budgets for all tasks; leave stdin delivery unchanged. Validation: make check passed all 68 tests, including delayed UTF-8 stdin/EOF, timeout and terminal-transition checks. All 556 current prompts fit 80000 characters; largest 77693. git diff --check passed. Foundation checks do not establish application readiness. Task: S00-013 Agent: auto-ci-cd-5f68f03dd33b Role: ci-cd Action: claim Status: review Task title: Bound oversized S21-006 runner context without truncating requirements
3.3 KiB
S00-013 executor evidence
Agent: auto-ci-cd-5f68f03dd33b; role: ci-cd; date: 2026-10-15. Result: bounded context preparation; pending independent QA review.
Input: assigned task/card/context, NFR-007, workflow/role instructions, S00-012/013 blocker reports, accepted scope ADR and S00-001/S00-011 dependency evidence/reviews/handoffs. Output: one runner context correction and focused regressions. No contracts or stack decisions changed.
Actual change and reasoning
S21-006 packet is 41660 characters and includes all 26 exact UI Kit sections. The old runner additionally embedded entire input documents, including the 11760-character stack ADR. Startup documents plus this packet and task inputs exceeded the 80000-character budget.
Runner now keeps assignment/startup instructions, exact packet, previous handoff and review evidence inline. Other task inputs are listed explicitly as required project reads before implementation/review, preserving their path annotations and missing-artifact blocker instruction. This also makes missing upstream inputs visible instead of silently omitting them. Project-root path checking remains enforced for referenced inputs. No text is sliced, no spec/packet is edited, and both packet and prompt budget guards remain active. The stdin transport implementation is untouched.
The S21-006 regression checks both executor and reviewer prompts, the whole exact packet, every verbatim UI Kit section, required input references and fresh-session/exit instructions. The all-current-prompts check now covers both actions for every task.
Actual commands and results
python3 scripts/tasks.py resume --agent auto-ci-cd-5f68f03dd33bandshow S00-013: exit 0; already assigned in_progress, no repeat claim.python3 scripts/tasks.py packet S00-013 --agent auto-ci-cd-5f68f03dd33b > /private/tmp/luma-S00-013-packet.md: exit 0.- Initial automation run of the first draft: exit 1, 23/24 tests passed; review prompt still exceeded budget when only upstream-annotated artifacts were deferred. Corrected by consistently referencing all supplementary task inputs, keeping startup instructions inline.
- Read-only Python enumeration using Runner with default config: exit 0; all 556 claim/review prompts for 278 tasks fit; largest S21-006 review 77693 characters, claim 77669 (agent auto-preview).
make check: exit 0, 68 tests passed. Final rerun after project-root guard: exit 0, 68 tests in 16.124s; validated 278 tasks, 22 stages, 221 requirements, 211 allocated Web requirements, revision 15. Includes delayed large UTF-8 stdin/EOF delivery, timeout, terminal-transition, orchestration, snapshot and CLI checks. Log: temporary /private/tmp/luma-S00-013-check.log.git diff --check: exit 0; source diff inspected, limited to scripts/runner.py and tests/tooling/test_automation.py. Initial state/tasks.json change was coordinator claim metadata.
Limits
These are foundation/harness checks, not app readiness or S00-002 acceptance. Agents must actually read the explicitly listed project inputs; references do not claim those artifacts were read automatically. Budget coverage describes current task/state artifacts; future larger evidence/handoffs still fail explicitly rather than truncate. No dependencies, credentials, production transports or immutable spec snapshots changed. Coordinator owns commit and the next fresh reviewer process.