Repository navigation
fix(subagent): bind experiment prerequisites and explain reruns - #267
Merged
Merged
Conversation
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
Prerequisite validation remains vulnerable to a launch race, supervised reports may omit provenance, and the PR metadata fails the documentation-category gate.
4 open findings
What changed in this PR
Adds durable prerequisite binding and rerun provenance to subagent execution.
Changes:
- Validates prerequisite receipts before and after resource admission.
- Requires rerun explanations and preserves provenance in records.
- Updates prompts, documentation, and tests.
| File | Description |
|---|---|
argus/tools/subagent/_cli.py |
Adds dependency and rerun CLI validation. |
argus/tools/subagent/_experiment_preflight.py |
Resolves and verifies prerequisite receipts. |
argus/tools/subagent/_direct_run.py |
Rechecks prerequisites before direct launch. |
argus/tools/subagent/_supervised_run.py |
Rechecks prerequisites before supervised launch. |
argus/tools/subagent/_registry.py |
Preserves provenance fields. |
argus/tools/subagent/_reporting.py |
Adds provenance to report generation. |
argus/roles/prompts/engineer.py |
Adds experiment prerequisite guidance. |
argus/roles/prompts/reviewer.py |
Adds independent-evidence guidance. |
docs/best-practices.md |
Documents prerequisites and reruns. |
tests/tools/test_subagent_prerequisites.py |
Tests validation and provenance behavior. |
tests/tools/test_subagent_durable_owner.py |
Adds durable CLI integration coverage. |
🧠 Review effort: Balanced
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+527
to
+529
| prerequisites = submitted_task.get("prerequisites") | ||
| if prerequisites: | ||
| resolve_prerequisites(prerequisites, expected_runs=prerequisites) |
Collaborator
Author
There was a problem hiding this comment.
Addressed in c3ddb04: the final receipt revalidation and process creation now share the submission admission lock in both execution modes. Added lock-boundary regression coverage and real cross-process exclusion coverage for the non-POSIX lock path. Later prerequisite reruns do not cancel already-started dependents.
Comment on lines
+352
to
+354
| for key in ("prerequisites", "previous_run_id", "rerun_reason"): | ||
| if key in task_data: | ||
| lines.append(f"- {key}: {json.dumps(task_data[key], ensure_ascii=False)}") |
Collaborator
Author
There was a problem hiding this comment.
Addressed in c3ddb04: both model-authored and deterministic reports append the attempt context directly from task data. Added a regression where the model intentionally omits all provenance fields.
Comment on lines
+834
to
+836
| prerequisites = submitted_task.get("prerequisites") | ||
| if prerequisites: | ||
| resolve_prerequisites(prerequisites, expected_runs=prerequisites) |
| shell quoting. Add `--bounded` when the objective is finite; without it the | ||
| worker keeps proposing follow-up work after the project is declared done. | ||
|
|
||
| ## Avoiding unproductive experiment reruns |
Reject failed or replaced prerequisites before command launch, require an explanation for identical resubmissions, and retain attempt context in reports. Apply evidence-independence guidance from the first review without adding a scheduler or replay policy. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Address independent review: hold the existing submission lock through the final prerequisite check and process creation, make its Windows fallback cross-process, and append run context deterministically to model-authored reports. Compress shared instructions within unchanged prompt budgets and rebuild the required release assets. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Restore operator-language, plan-authority and checkpoint boundaries lost during prompt compression. Reuse the four failing regression checks and keep every evidence form within the unchanged 5700-character budget. Rebuild shipped release artifacts. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace repeated prefix scans with incremental UTF-8 fragments. Retain exact byte-limit, CRLF, EOF, queue and cancellation behavior, and reproduce the original API overflow work before the fix. Preserve upstream Reviewer contracts without raising the policy budget. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
aHappend
force-pushed
the
fix/experiment-evidence-discipline
branch
from
October 10, 2026 11:28
50014ba to
dde99e2
Compare
Reuse operator settings and round controls for progress-first, standard and strict evidence policies. Preserve independent acceptance and existing prompt budgets. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Share the design-symptom rule, retain the fixed 5400-character Engineer budget for every policy, and migrate legacy prompt assertions to the shared contract without weakening behavior checks. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Change
--depends-ontasks to successful existing run receipts. Recheck them at the real launch boundary, including after resource admission and with semantic preflight disabled.--rerun-reasonin terminal records and reports. A reason string is not proof that the inputs changed.ARGUS_SKILL_EVIDENCE_POLICY=progress-first|standard|strictthrough the existing persisted operator settings and public configuration endpoint. Invalid values raise errors; explicit environment values take precedence.progress-first, reuse the existing round policy: cap enabled no-progress streak thresholds at two and enabled soft-round thresholds at three. Preserve disabled guards, explicit mission settings and individual operator overrides. Real forward progress continues through the existing guard logic; no new retry loop or evidence ledger.The evidence profiles control relevant proof depth and existing continuation behavior. They are not a full provider-stream auditing implementation and do not retrospectively accept failed executions.
Current integration
Rebased onto
devat60287bcc8a1b51e34cbf09dfe5404e88fa1d5180, then pushed997e45ae4135a804233886eca3f2b52774126941.The standard release builder regenerated the manifest and Web/TUI artifacts for
0.1.8+b3af822cd5f0d584. Artifact consistency passes.Current targeted validation: 270 passed, covering public settings persistence and rejection, all role consumers, real SkillLoop execution with native memory-backend review actions, progress/no-progress behavior, configuration precedence, and unchanged Engineer/Planner/Reviewer prompt budgets. Full CI on the preceding head exposed stale text assertions and a 14-character ordinary-Engineer budget overrun; the follow-up compresses shared rules and covers all three policies without increasing any budget. Scoped Ruff and diff checks pass. The current-head CI status is in the Checks tab.
The earlier prerequisite/Reviewer/TypeScript changes were validated on
dde99e21: 458 Python tests passed with 16 existing platform/opt-in skips, all 235 Node tests and TypeScript typecheck passed, and Linux plus Ubuntu/macOS/Windows CI passed. Those are historical results, not a claim about the new head's pending full CI.Documentation:
docs/best-practices.mdcovers evidence profiles, prerequisite admission and rerun limits;docs/typescript-migration.mdcovers incremental buffering.Not merged. Existing running installations do not adopt a pushed branch automatically.