Repository navigation
M11: autonomous by default; only one-way doors go to the human - #122
Merged
Merged
Conversation
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FeZyj9c1mTZ8GsF1SyvzYR
…rant; a two-way ask takes its recommended answer Only one-way doors go to the human (generation is cheap: a reversible choice is redone). - events authority: --mode defaults to autonomous (AGENTIC_AUTONOMY_MODE=reviewed opts in) - _approval: a sprint planned from an intent whose words are exactly a captured human prompt is granted autonomy with no second line; untyped words, a short reply or a re-pointed plan still ask - intent: in autonomous mode an ask with a recommended answer and no door is taken and recorded (intent.normalized taken=[...]); a door or no recommendation still asks Tests: plan-approval 22 (new check red on master), intent-contract 16 (new check red on master). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FeZyj9c1mTZ8GsF1SyvzYR
…-question init; docs aligned Only one-way doors go to the human; the door-classifier decides which, before execution. - agents/door-classifier.md (planning route) + gate-door-plan (SubagentStop): every task tagged - mission doors tag/plan: tags only add; open one-way doors listed together; gate-dispatch parks only those tasks - mission doors reclassify: ALOOP_DOORS_TWO_WAY plus the human's typed 'two-way door:<id> <code>' (ledger row door.reclassified, M9.17 provenance); walls and question-only doors refused - init skill applies the defaults with no questions (--ask restores them) - OVERVIEW, README, plugin README, SKILL, CLAUDE.md, site, ROADMAP/TRACKING aligned Tests: mission-door-plan 6, mission-doors-reclassify 5, make test-hosts green, eval-validate, check-tracking. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FeZyj9c1mTZ8GsF1SyvzYR
adeelahmad
marked this pull request as ready for review
November 19, 2026 13:35
… a typed prompt can be the grant - _approval.said: the whole prompt or its first line (after an optional slash command) - e2e-live.sh: no approval line; the prompt's first line is the intent - intake persona: --raw is the human's words verbatim Tests: plan-approval 22, e2e-live-hosts 16. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FeZyj9c1mTZ8GsF1SyvzYR
…rant, 12/12, $0.82 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FeZyj9c1mTZ8GsF1SyvzYR
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.
Requested by Adeel · project thread
What
Core rule (owner, 2026-11-19): generation is cheap, so only one-way-door decisions go to the human. ROADMAP §4.14 has the audit (A1–A12, every place a run waited for the human, with file:line) and the design.
Implemented:
events authoritywithout--mode); reviewed is opt-in (AGENTIC_AUTONOMY_MODE=reviewed)./aladdin) gets autonomy with noautonomous SN <code>line. An untyped paraphrase, a short reply, or a plan re-pointed at another intent still asks / loses the grant.askwith the intake'srecommendedanswer and nodooris taken and recorded (intent.normalized taken=[…]); one naming a one-way door, or with no recommendation, still asks.door-classifierrole (planning route) tags each planned task one-way (mission doors tag T --door ID) or two-way.mission doors plan SNlists the open one-way doors together;gate-door-plan(SubagentStop) blocks the classifier's stop while a task is untagged or a wall is planned;gate-dispatchparks only tasks with an unauthorized door. Tags only add, and the PreToolUse door gate stays the backstop.ALOOP_DOORS_TWO_WAY=<door,…>makes a door two-way only while adoor.reclassifiedledger row holds. That row counts only with the human's typedtwo-way door:<id> <code>line (M9.17 provenance), so an agent can't forge it. Walls and question-only doors are refused, andrevokecloses it.--askbrings the questions back).No gate's deterministic checks were weakened. Every grant still rests on a captured human prompt.
Call: an untagged task is treated as two-way at dispatch (the door gate is the backstop), so older plans and fixtures keep running. The classifier's own stop gate is what requires every tag.
Next: M11.5 sprint chaining, M11.3 plan judge, M11.7, M11.8, M11.9 live run.
Evidence
plan-approval.test.py22,intent-contract.test.py16.mission-door-plan.test.py6,mission-doors-reclassify.test.py5.make pycheck test-hostsgreen,make eval-validate(18 suites),check-tracking(111 requirements).🤖 Generated with Claude Code
https://claude.ai/code/session_01FeZyj9c1mTZ8GsF1SyvzYR