@davidorex/pi-jit-agents 0.14.6 → 0.28.0

package/templates/task-worker/task.md

@@ -0,0 +1,52 @@
+## Task: {{ task.id }}
+
+### Description
+{{ task.description }}
+
+{% if task.phase %}
+### Phase
+{{ task.phase }}
+{% endif %}
+
+### Files
+{% for f in task.files %}
+- `{{ f }}`
+{% endfor %}
+
+### Acceptance Criteria
+{% for criterion in task.acceptance_criteria %}
+- [ ] {{ criterion }}
+{% endfor %}
+
+{% if task.depends_on %}
+### Dependencies (completed)
+{% for dep in task.depends_on %}
+- {{ dep }}
+{% endfor %}
+{% endif %}
+
+{% if task.notes %}
+### Notes
+{{ task.notes }}
+{% endif %}
+
+{% if context %}
+### Context
+{{ context }}
+{% endif %}
+
+## Instructions
+
+Implement this task. Follow these steps:
+
+1. **Read the target files** listed above to understand the current state
+2. **Implement changes** that satisfy each acceptance criterion
+3. **Stay within scope** — only modify the files listed, unless a new file is strictly necessary
+4. **Run tests** after implementation to verify nothing is broken
+5. **Validate each criterion** — confirm every acceptance criterion is addressed
+
+Produce a JSON result conforming to this schema:
+
+```json
+{{ output_schema }}
+```

package/templates/verifier/task.md

@@ -0,0 +1,57 @@
+You are verifying the output of a prior step against declared success criteria.
+
+## Success Criteria
+
+{% for criterion in criteria %}
+- {{ criterion }}
+{% endfor %}
+
+## Step Output to Verify
+
+```json
+{{ step_output | dump(2) }}
+```
+
+{% if files_to_check %}
+## Files to Check
+
+{% for f in files_to_check %}
+- `{{ f }}`
+{% endfor %}
+{% endif %}
+
+## Instructions
+
+For each criterion, choose the appropriate verification method and execute it:
+
+1. **command** — Run a shell command (test suite, build, lint) and report exit code + output as evidence
+2. **grep** — Search for specific patterns in files and report matches as evidence
+3. **inspect** — Read files and assess their content (use your judgment)
+4. **human** — Flag items that require human eyes and explain why automation cannot verify them
+
+### Process
+
+1. Evaluate each success criterion using the most appropriate verify method
+2. For `command` criteria: actually run the command and capture the output
+3. For `grep` criteria: search the relevant files and report what you find
+4. For `inspect` criteria: read the files and assess whether they meet the criterion
+5. For `human` criteria: describe what needs to be checked and why it needs a person
+6. Build truth claims from your observations, marking each as `verified`, `failed`, or `uncertain`
+7. Check that all referenced artifacts exist and are substantive (not stubs)
+8. Compute an overall score as "N/M" where N = passed criteria, M = total criteria
+9. Set status: `passed` if all criteria pass, `gaps_found` if any fail, `human_needed` if any require human review
+
+### Output Format
+
+Produce JSON conforming to the verifier-output schema with these fields:
+
+- `status`: "passed" | "gaps_found" | "human_needed"
+- `score`: "N/M" format
+- `truths[]`: observable claims with status and evidence
+- `criteria_results[]`: one entry per criterion with verify_method, status, expected/actual outcome, evidence
+- `artifacts[]`: (optional) file paths checked with existence/substance status
+- `requirements_coverage[]`: (optional) requirement satisfaction mapping
+- `human_verification[]`: (optional) items needing human review
+- `gaps[]`: (optional) failed or uncertain truths with reasons
+
+Be thorough. Run actual commands. Read actual files. Report what you observe, not what you assume.