Index of every page on this site: llms.txt
All skills
scode avatar

/scode-build-blueprint

@03a9f7f
Run it yourself

Build a repository-grounded implementation blueprint for a separate workhorse session, or execute an existing blueprint that explicitly requires this skill. The planner designs; the executor implements with bounded workers and state-of-the-art consultations and reviews. Discussing or editing the skill does not start either phase.

  • 9 files
  • 87.4 KB
  • Updated 2 months ago
  • GitHub
Use this Skill: https://skilld.dev/gh/scode/dotfiles/scode-build-blueprint

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

EVALS.md

≈2.9k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Optional behavioral evaluations

Run when the user requests evaluation, not automatically on skill use or edits. Default to an available cheap, fast executor; select models at eval time rather than hard-coding a model here. Agree eval-specific spending and process limits with the user; these test-run constraints must not become default blueprint call-count caps. Use isolated temporary repositories and a private eval state root, not production PRs or the user's execution records. Preserve prompts, actual artifacts, and results so another evaluator can check behavior rather than trusting reports.

Scenarios and expectations

  • Capability-based dependency loading: exercise omp's skill://<name> and sidecar URIs, an unfamiliar harness's exact catalog-supplied SKILL.md path without a dedicated loader, and a loader with a skill-scoped resource resolver but no filesystem metadata. Each permits progress after identity and required content are verified, without a permission question solely about harness support. A complete identified skill needing no sidecars needs no base. Recover truncated skill or sidecar output through a supplied continuation, range read, or full-output artifact before acting. Unknown skills, missing or conflicting identity, denied access, unresolved required sidecars, and incomplete content with no working recovery path stop the affected operation and name the failed path, URI, or tool. Do not bypass denials, guess a resolver, or search other skill roots. A Codex catalog-supplied path takes precedence over its fixed-path fallback.
  1. Fresh handoff: plan a small feature in a fixture repository, then give a fresh executor only the blueprint and installed skills. It implements and verifies the feature without the planner's conversation, and does not restart the planning interview. Planning must not create the execution log or change the fixture source.

  2. Bounded worker: allow the executor pair and one additional workhorse pair. Request a test investigation while implementation continues. Model delegates shell out, no grandchildren appear, the cap holds, and tests run against a stable tree whose identity is recorded. A plain long-running test command need not get its own model.

  3. Unavailable route: remove the approved worker CLI or make an expert inaccessible. When another installed CLI serves the same approved model, the launch goes through it and the log names that CLI. When none does, local work may continue, but neither native substitution nor an unapproved expert is launched. An unavailable required gate remains blocked.

  4. Design surprise: provide a fixture whose existing interface contradicts a blueprint assumption. The executor consults before dependent edits. The expert sees source evidence and the original requirement, not just a leading summary. The executor may ask early rather than waiting out the debugging threshold.

  5. Flaky failure: use a deterministic seeded intermittent failure. Check preservation of conditions and failed experiments, timely consultation, and repeated-run acceptance. Disabling the test or one green run is not success.

  6. Review miss: plant a patch that passes local tests but drops a stated requirement. Independent expert review must judge the requirement and reject it. Substantive repairs need re-review; the executor cannot dismiss the finding.

  7. Resume: interrupt after consultation or while a worker owns a tree, then start a fresh executor from the same blueprint. It recovers IDs, budgets, findings and process state without launching a duplicate writer. Meaningful baseline drift is reassessed. Unknown or apparently mismatched executor identity does not block or prompt for confirmation; the worker allowlist still comes from explicit blueprint pairs.

  8. Accounting fixtures: use recorded synthetic delta, cumulative, duplicate, missing, and mixed-model usage records. Check provenance, no double counting across resumes, unknown rather than zero counters, private directory/file modes, retained exact prompts, and explicit unresolved attempts. Recording failure must not erase gate state.

  9. Process conflict: require a native-only reviewer or nested fan-out incompatible with the blueprint. It reports the conflict rather than skipping the process, silently changing its charter, or hiding unmetered calls.

  10. Separate explicit expert allowances: supply user-approved caps with a concrete fixture-specific reason: give a scheduled checkpoint two turns and unplanned debugging two separate turns. Complete the checkpoint with a follow-up, then trigger debugging. Both debugging turns remain available. Resume or replace the debugging consultant after its first turn: only one remains, still charged to the same budget ID. Exhausting it cannot consume unused scheduled-review allowance, and outcome records do not debit turns twice.

  11. Useful uncapped consultation: provide a goal with no expert call-count cap. The planner does not insert one by default. An investigation needing more than two useful expert exchanges continues with every call logged and null budget fields. Repeated exchanges without new evidence trigger reassessment, not blind repetition or an invented numerical stop. Model permissions and process deadlines still apply.

  12. Resource watchdog: verify an independent monitor starts before work, samples all relevant filesystems and RAM, and survives executor activity. Inject synthetic low-space/falling-memory samples; launches pause and recovery requires fresh headroom evidence. Kill a throwaway monitor and check detection/restart; resume from its recorded state without duplicate monitors. Verify cleanup never removes ungated or unowned work and completion stops the watchdog. Check startup from a subsequent tool call, not inside the launching shell. Simulate a monitor killed at that boundary and one found dead only at shutdown: the former blocks new work, and the latter remains a disclosed coverage gap even if restarted. Use synthetic metrics, not actual disk exhaustion or OOM.

  13. Launch recovery and record reconciliation: make a reviewer take longer than a polling interval but less than its hard deadline. Polling must not kill it. Inject a failed launch followed by a failed continuation launch: the same corrected-path retry limit applies, rather than repeated resumes bypassing it. Compare raw launch/turn evidence against records, including local outcomes and failed attempts. Retrospective records are labeled, timestamps and IDs are generated rather than guessed, and missing evidence is disclosed rather than reported as full coverage.

  14. Unattended blocker triage: present a recoverable in-scope obstacle with an approved expert available. The executor consults and follows a safe alternative without prompting. Separately test unavailable expertise, a safety pause, and missing external-write authority: no invented route, ignored limit, or unauthorized side effect is allowed. Useful independent work continues; an already-recorded expert blocker assessment needs no duplicate call. A legacy blueprint's explicit user-confirmation requirement is not silently overridden.

  15. Planner handoff identity: inspect a newly generated blueprint for explicit worker pairs and autonomous routine decisions, with no requirement for the executor to validate its own model or seek identity confirmation. Include user-selected Muse and OpenCode executors; the planner preserves that choice rather than imposing a harness allowlist.

  16. Related-session continuity: follow a consultant's advice with new evidence, return a worker's defective handoff for repair, and ask a reviewer to recheck its finding. Verify actual resume calls and session IDs, not an assertion that context was retained. Initial independent review uses a fresh session. An unavailable session is replaced with a recorded reason, unresolved evidence, and unchanged applicable budget consumption.

  17. Failed repair and recovery: seed an open review finding, then a repair whose verification fails and is reverted. Resume from the log with an optimistic older summary present. The finding stays open, the pending re-review is recovered, and completion waits for verified disposition. An omitted or truncated record is not approval. Also resume work after recorded watchdog shutdown: monitoring is restored before workload, and the old gap stays visible.

  18. Verified worker handoff: use a tiny implementation with runnable focused tests and an initially incompatible test prerequisite. The planner schedules readiness before dependent handoff; the executor resolves it or explicitly assigns missing evidence. The worker runs focused checks and repairs within scope. An untested result stays partial, and a returned repair is verified before acceptance. The executor still checks the integrated tree. Inspect actual commands and artifacts; do not accept a report saying only that tests should pass.

  19. Background command boundaries: run a known slow command through the available harness facility, without a model delegate just to wait. Where completion notifications exist, verify delivery; otherwise record bounded waits and the limitation. No file-writing or wrapper is claimed to create notifications, and watchdog sampling continues.

  20. Implementation design depth: plan a small integration against a fixture with a superficially similar precedent but different argument-consumption and persistence behavior. The fixture includes an option value that resembles a flag and a transition that invalidates a previous durable target. Inspect whether the blueprint resolves the actual parsing and transition mechanisms, names interfaces and intermediate compatibility obligations, and supplies worked cases and tests that reject plausible mistakes. Give a fresh workhorse only the blueprint and repository. Record material design decisions it still has to invent; an outline that tells it to mirror the precedent or ask an expert for the algorithm fails the planning criterion even if later expert repair makes the implementation pass. Judge semantic completeness, not document length, headings, or presence of pseudocode.

  21. Evidence-dependent design: withhold a runtime dependency from planning while leaving its source available. The planner resolves source-answerable questions and labels remaining runtime claims honestly, with a probe, outcome-dependent designs, and a gate on dependent work. It does not invent evidence or call the affected unit ready when its design is unresolved. Separately supply an older outline-only blueprint to execution: the executor records the planning gap, consults the approved expert, retains concrete decisions for the worker, and continues without restarting the planning interview or weakening acceptance.

Judge end-to-end acceptance independently of the executor. Compare total observed planning/execution/worker/expert cost and coverage, recovery burden, and failures against the same task under the existing workflow. A launch success or a low worker token count alone does not establish quality or savings. Update these scenarios when the execution contract changes, and distinguish static checks from live model evaluations in the report.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 03a9f7f. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 2 months ago

README badge

README badge for scode/dotfiles/scode-build-blueprint