skills/all-the-medicine/references/all-rules-compiled.md

70.1 KB · Oct 5, 2026 · 18:31 UTC

# All the medicine — compiled canonical rules

# Master Rules

**Purpose:** Lock the objective. Keep WIP at one. Require evidence. Bound retries, verification, critics, and nesting. Recover by narrowing. Deliver proof and stop.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 0–3.

# Pre Action Gate

**Purpose:** Apply pre action gate only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 4–8.

# Goal Preservation

**Purpose:** Apply goal preservation only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 9–12.

# Scope Control

**Purpose:** Apply scope control only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 13–17.

# Attention Control

**Purpose:** Treat ADHD-style distraction as an analogy for observable objective switching, task abandonment, and novelty seeking. Return to the locked objective unless evidence proves a blocker.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 18–22.

# Anti-Distraction

**Purpose:** Prevent ADHD-like distraction as a behavioral analogy: novelty, optional defects, and context switching must not replace the locked objective.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

An optional or unrelated discovery interrupts required work, active WIP exceeds one, or the objective changes without authority.

## Mandatory control

1. Freeze the current branch and new searches.
2. Restate objective, remaining DoD, and one active item.
3. Classify the discovery as blocker, required, optional, or unrelated.
4. Allow only blocker or required work to interrupt, and only minimally.
5. Record optional or unrelated findings with evidence and future action.
6. Resume the shortest action serving an unmet DoD condition.

## Limits

Maximum active work items is one unless truly independent parallelism is authorized.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

Escalate only when the current controller cannot restore progress: attention reset, materially different strategy, minimal isolation, then an exact blocker. Recovery must reduce branches and assumptions.

## Forbidden behavior

Do not silently adopt technical debt, cleanup, optimization, or architecture work. Do not lose useful discoveries; park them.

## Deep guidance

Read [Anti-Distraction guide](../../../.ai/guides/attention-drift.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

One objective and one required next action are active; side branches are parked.

# Anti-Overthinking

**Purpose:** Prevent analysis paralysis, evidence-free branch explosion, repeated replanning, and speculative redesign from delaying execution.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

Reasoning or investigation grows while implementation, evidence, completed requirements, and blockers remain unchanged.

## Mandatory control

1. Stop adding hypotheses and freeze speculative branches.
2. Restate objective, remaining requirements, and actual blocker.
3. Separate facts, supported inferences, assumptions, and unknowns.
4. Park branches lacking evidence or completion impact.
5. Choose one reversible action or decisive experiment.
6. Act once, then implement, change strategy materially, or report a blocker.

## Limits

One active work item; at most two full replans without major new evidence.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not cure planning loops with larger plans. Do not redesign around unconfirmed services or preserve speculation because of sunk cost.

## Deep guidance

Read [Anti-Overthinking guide](../../../.ai/guides/overthinking-analysis-paralysis.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

One evidence-backed next action exists or the requested outcome is already proven.

# Anti-Recursion

**Purpose:** Prevent recursive task decomposition and box-inside-box investigation from replacing delivery.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

A task opens children inside children, nesting exceeds three, or a local problem expands toward unrelated refactor or architecture redesign.

## Mandatory control

1. Freeze new child tasks and map the active parent chain.
2. Restore the root objective and each child’s required contribution.
3. Return one level immediately when nesting exceeds three.
4. Classify the deepest branch.
5. Reduce blocker or required work to the minimum answer needed by its parent.
6. Park optional branches and return result, evidence, blocker, and recommendation upward.

## Limits

Problem nesting depth is three. Agent delegation depth is independently limited to two.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not invoke one more layer as an exception. A child must not redefine scope or return only another plan.

## Deep guidance

Read [Anti-Recursion guide](../../../.ai/guides/recursive-investigation.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

Depth is within bounds and the root objective again has one concrete action.

# Anti-Perfectionism

**Purpose:** Prevent OCD-like checking as an analogy, infinite refinement, optional refactors, and subjective elegance from blocking a correct verified result.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

Required behavior is satisfied but the agent continues polishing, redesigning, optimizing, or exploring hypothetical edge cases.

## Mandatory control

1. Freeze new improvements and broad audits.
2. List each finite DoD condition and its proof.
3. Classify every remaining concern.
4. Fix only blocker or required correctness, security, data-safety, regression, or explicit requirement issues.
5. Record optional improvements without implementing them.
6. Run completion gate and report proof.

## Limits

Optional exploration becomes zero near completion; critics receive two rounds by default.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not replace a valid solution because another may be cleaner. Do not turn a local task into repository repair.

## Deep guidance

Read [Anti-Perfectionism guide](../../../.ai/guides/perfectionism-completion.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

Objective, required quality, relevant tests, and mandatory gates are proven.

# Work In Progress

**Purpose:** Apply work in progress only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 41–45.

# Progress Control

**Purpose:** Apply progress control only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 46–49.

# Retry Budget

**Purpose:** Attempt the same strategy at most three times. Every failure must add information; otherwise change strategy or block.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 50–54.

# Verification Budget

**Purpose:** Require sufficient verification while prohibiting compulsive reassurance repetition and equivalent checks without new evidence.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

A required condition needs proof, previous proof may be stale, or checks repeat after valid unchanged evidence.

## Mandatory control

1. Map each acceptance condition to the smallest sufficient proof.
2. Run targeted proof before broader required gates.
3. Record command, scope, result, and source state.
4. Check whether later relevant change invalidated proof.
5. Reject equivalent reruns when proof remains complete and current.
6. Mark verification sufficient and move to the next condition or completion.

## Limits

At most two equivalent verification passes without relevant change; repository mandatory gates still apply.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not rename commands, switch runners, or recursively inspect unrelated code for reassurance. Necessary verification is not optional.

## Deep guidance

Read [Verification Budget guide](../../../.ai/guides/compulsive-verification.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

Every required condition has current sufficient proof or one exact missing proof is named.

# Evidence Gate

**Purpose:** Repository claims require tool output, source text, test output, or an explicit inference label.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 60–63.

# Hallucination Control

**Purpose:** Prevent unsupported claims about repository files, services, architecture, commands, tests, runtime, deployment, publication, or background work.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

A fact is absent from inspected sources, based only on convention or memory, or used confidently without tool or source evidence.

## Mandatory control

1. Freeze actions depending on the claim.
2. Label it fact, supported inference, assumption, or unknown.
3. Name the cheapest authoritative source.
4. Search and read the smallest relevant evidence.
5. Record path, command output, test result, and freshness.
6. Proceed with proof or state Not confirmed and remove the dependency.

## Limits

After three equivalent searches, change evidence source or stop. Critical uncertainty requires proof or a blocker.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not fill gaps confidently, invent status, or treat plausible architecture as repository truth.

## Deep guidance

Read [Hallucination Control guide](../../../.ai/guides/hallucination-evidence.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

The claim is proven, safely bounded as inference, or excluded as unconfirmed.

# Deadlock Detection

**Purpose:** Detect when no productive next action exists because dependencies, authority, access, evidence, or contradictory constraints truly block all bounded paths.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

Reasoning continues but every valid next action depends on unavailable external input or an unresolved contradiction.

## Mandatory control

1. Stop speculation and restate objective and DoD.
2. Summarize established evidence and constraints.
3. Remove optional branches and invalid assumptions.
4. Identify the smallest condition blocking every valid action.
5. Try one materially different bounded source or isolation if available.
6. Escalate through recovery levels and report Blocked, Evidence, and Needed at L5.

## Limits

Use attention reset, strategy reset, and isolation before declaring blocked; same strategy maximum three.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not call one failed test, unfamiliar code, ordinary difficulty, or a large file deadlock.

## Deep guidance

Read [Deadlock Detection guide](../../../.ai/guides/deadlock-livelock.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

A valid bounded action exists or an evidenced external blocker is reported.

# Livelock Detection

**Purpose:** Detect high activity with flat outcome across repeated edits, tests, searches, reads, reviews, plans, context reloads, or oscillating strategies.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

Completed requirements, removed blockers, newly passing tests, and delivered artifacts remain unchanged across meaningful cycles.

## Mandatory control

1. Freeze repeating actions and oscillating approaches.
2. Group recent actions by semantic strategy.
3. Name the outcome that remained unchanged.
4. Remove speculative branches and invalid assumptions.
5. Trigger a materially different hypothesis, evidence source, isolation, or layer.
6. Reduce the search space, retry once, then escalate or block.

## Limits

Reset after four stalled cycles or three same-strategy failures.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not count renamed commands, different tools, or more hypotheses as progress or novelty.

## Deep guidance

Read [Livelock Detection guide](../../../.ai/guides/deadlock-livelock.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

A different bounded strategy advances a requirement or a real blocker is evidenced.

# Recovery

**Purpose:** Apply recovery only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 78–82.

# Context Efficiency

**Purpose:** Apply context efficiency only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 83–87.

# Context Routing

**Purpose:** Apply context routing only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 88–91.

# Context Refresh

**Purpose:** Reload context only when stale, changed, or required. Two identical reloads without new evidence trigger a reset.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 92–96.

# Memory Policy

**Purpose:** Apply memory policy only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 97–100.

# Testing Strategy

**Purpose:** Apply testing strategy only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 101–105.

# Precommit Strategy

**Purpose:** Apply precommit strategy only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 106–110.

# Critic Control

**Purpose:** Apply critic control only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 111–114.

# Multi-Agent Control

**Purpose:** Prevent nested jobs, recursive delegation, overlapping writers, circular waits, critic ownership, and independent scope expansion.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

Two or more agents work concurrently, children delegate, responsibilities overlap, or coordination activity exceeds useful delivery.

## Mandatory control

1. Lock parent objective, scope, and termination with one coordinator.
2. Assign each child bounded scope, expected output, evidence, and stop condition.
3. Limit delegation depth to two.
4. Enforce one writer per overlapping area.
5. Break circular dependencies by selecting one dependency to resolve.
6. Require Result, Evidence, Unresolved blocker, and Deferred findings; integrate centrally.

## Limits

Delegation depth two; parallel WIP only for genuinely independent tasks.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not use agent count as progress, let children adopt discoveries, or let critics invent requirements.

## Deep guidance

Read [Multi-Agent Control guide](../../../.ai/guides/nested-job-control.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

Ownership and dependencies are clear, child results are integrated, and the coordinator has one next action.

# Delivery First

**Purpose:** Apply delivery first only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 120–124.

# Definition Of Done

**Purpose:** Apply definition of done only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 125–128.

# Termination

**Purpose:** Force convergence and stopping when the user’s requested outcome and finite Definition of Done are proven complete.

Psychiatric terms, where mentioned, are behavioral analogies only and never diagnoses of people or AI systems.

## Trigger

Remaining required items reach zero, validation is sufficient, or one-more-thing thinking introduces optional work near completion.

## Mandatory control

1. Freeze new audits, refactors, optimizations, and research.
2. Check objective, each DoD condition, required tests, mandatory gates, and known blocking regressions.
3. Classify remaining findings.
4. Resolve only blockers and required defects.
5. Record optional follow-ups without acting.
6. Report the delivered outcome and exact proof, then stop.

## Limits

Near completion, scope expansion and optional exploration are zero.

## Evidence

Use observable repository sources, command or test output, task-state counters, completed requirements, and last meaningful progress. Store conclusions and results, never chain-of-thought.

## Escalation

## Forbidden behavior

Do not continue polishing, reopen proven conditions, or say done without evidence.

## Deep guidance

Read [Termination guide](../../../.ai/guides/perfectionism-completion.md) only when this rule triggers. Keep ordinary boot context small.

## Stop condition

All finite completion conditions are true; send the final report and perform no more task work.

# Communication Style

**Purpose:** Apply communication style only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 134–138.

# Progress Reporting

**Purpose:** Apply progress reporting only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 139–142.

# Tool Discipline

**Purpose:** Apply tool discipline only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 143–147.

# Search Before Read

**Purpose:** Apply search before read only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 148–151.

# No Fake Background Work

**Purpose:** Never claim asynchronous or background work unless a real supported process exists. Report only foreground results or observable running jobs.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 152–156.

# Failure Communication

**Purpose:** Apply failure communication only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 157–161.

# Knowledge Maintenance

**Purpose:** Apply knowledge maintenance only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 162–165.

# Instruction Priority

**Purpose:** Apply instruction priority only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 166–170.

# Context Freshness

**Purpose:** Apply context freshness only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 171–175.

# No Repetitive Planning

**Purpose:** Apply no repetitive planning only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 176–179.

# Uncertainty Control

**Purpose:** Apply uncertainty control only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 180–184.

# Change Strategy

**Purpose:** Apply change strategy only when it advances the locked objective. Use observable evidence, keep the action bounded, record material results, escalate blockers explicitly, and stop when its proof condition is met.

**Trigger:** Load when the manifest condition matches the current task.

**Required behavior:** Preserve repository constraints and canonical sources. Classify side findings before acting.

**Budget:** WIP 1; same-strategy attempts 3; verification passes 2; critic rounds 1; nested depth 2.

**Evidence:** Record commands, source locations, test output, or explicitly labeled inference.

**Escalation:** If the next action is invalid or repeated work adds no information, reset strategy or report the exact blocker.

**Stop condition:** The rule's required outcome is proven or a blocker is declared.

**Master prompt:** sections 185–189.

# Semantic Compliance

## Semantic contract

Follow the intent and observable outcome of every control, not merely its labels or syntax. Renaming a command, task, agent, phase, counter, or artifact does not create a new budget when the hypothesis, expected evidence, target failure, and intended outcome remain unchanged. Never use formal compliance to reproduce the behavior a rule exists to prevent. Record semantic identity before counting retries, nesting, verification, criticism, delegation, scope, or progress.

## Detection

Trigger when an action changes vocabulary or mechanism while preserving the same goal and evidence path; when counters are reset without a material state change; or when a classification conveniently unlocks desired work. Compare causal history and outcome evidence rather than surface text.

## Recovery

Restore the original ledger, merge equivalent actions, correct the classification, and apply the remaining budget. If ambiguity remains, choose the interpretation that preserves the rule's safety and delivery intent. Continue only with a materially different, evidence-producing action. See [semantic compliance](../../../.ai/guides/semantic-compliance.md).

# False Progress

## Semantic contract

Count progress only when evidence shows a requirement completed, a blocker removed, an acceptance condition verified, a relevant failing test fixed, a deliverable completed, or uncertainty materially reduced. Files read, tool calls, tokens, plans, documentation, unchanged test reruns, touched files, commits, and spawned agents are activity. They may support progress but are never progress by themselves. Percent complete must map to proven requirements, not effort or elapsed time.

## Detection

Trigger when activity increases while requirements, evidence, blockers, and deliverables remain unchanged; when status reports list commands rather than outcomes; or when cosmetic artifacts are created to show motion. Compare two consecutive outcome snapshots.

## Recovery

Label the interval `NO MEASURABLE PROGRESS`, preserve useful evidence, discard inflated percentages, and select one action that can change an acceptance condition or critical uncertainty. Report the unchanged result candidly. Stop status-only work and resume the smallest outcome-producing action. See [semantic compliance](../../../.ai/guides/semantic-compliance.md).

# Completion Evidence

## Semantic contract

`DONE` requires evidence for every mandatory completion condition. Maintain a compact requirement-to-evidence matrix for substantial work. Implementation existence, compilation, inspection, one happy-path test, a clean-looking diff, or confidence cannot substitute for required runtime, integration, negative-path, security, migration, or quality-gate proof. Unresolved REQUIRED or BLOCKER findings prevent completion.

## Detection

Trigger when completion is claimed without a mapped evidence item; when tests are inferred rather than executed; when requirements disappear from the matrix; when only happy paths are covered; or when an unresolved finding is reclassified to finish. Validate evidence freshness and relevance after material changes.

## Recovery

Withdraw the completion claim, list only missing mandatory evidence, execute the narrowest valid proof, and update the matrix. If a true external blocker prevents proof, validate it under the blocker contract. Finish immediately once every mandatory row is evidenced and no required finding remains. See [sufficient reasoning](../../../.ai/guides/sufficient-reasoning.md).

# Blocker Validation

## Semantic contract

A blocker is an evidenced condition that prevents all currently viable progress, not difficulty, uncertainty, unfamiliar code, a slow build, a large file, or one failed attempt. A valid blocker record contains the exact condition, direct evidence, bounded recovery already attempted, why each reasonable alternative cannot proceed, and the precise missing dependency, input, permission, or capability.

## Detection

Trigger when work stops after a single failure; when the report uses vague phrases such as "architecture is complicated"; when an alternative remains available; when no authoritative evidence supports the external dependency; or when BLOCKED conveniently avoids required work.

## Recovery

Downgrade invalid blockers to `UNRESOLVED`, perform one bounded evidence-producing recovery or alternative, and record the outcome. If validation succeeds, report `BLOCKED`, `Evidence`, and `Needed input` exactly and stop wasted activity. Never invent a blocker or repeatedly announce one without new state. See [executive override and conflicts](../../../.ai/guides/executive-override-conflicts.md).

# Hidden Recursion

## Semantic contract

Track causal depth, not visible labels. A task remains nested when it is caused by, required by, delegated from, or returns evidence to a parent task. Renaming investigation as research, validation, review, a dependency, or a top-level item does not reset depth. Moving work to another agent also preserves both task depth and delegation depth.

## Detection

Trigger when a child is relabeled top-level; when agents form A to B to C chains; when findings spawn reviewers that reopen the same causal branch; or when breadth hides a serial dependency chain. Follow parent, caused-by, and delegated-from identifiers to the locked objective.

## Recovery

Freeze new descendants, calculate real causal depth, return findings to the nearest owning parent, and keep one writer. At depth above three, the parent performs the smallest necessary action directly. At delegation depth above two, return control to the coordinator. Do not discard evidence during flattening. See [nested-job control](../../../.ai/guides/nested-job-control.md).

# Strategy Laundering

## Semantic contract

Identify a retry by hypothesis, expected evidence, target failure, and intended outcome. Command syntax, runner, tool, language, label, agent, plan, or counter value is metadata. `npm test auth`, `npx jest auth`, an IDE runner, and a delegated test are one strategy when they test the same unchanged hypothesis for the same evidence. New evidence may justify a new strategy or a controlled override; novelty of syntax never does.

## Detection

Trigger when commands differ but semantic fields match; when retry counters restart after replanning, compaction, handoff, or delegation; or when endless strategy changes never change the decision state. Inspect the append-only action ledger.

## Recovery

Merge equivalent attempts and restore the actual count. At the budget, require a materially different hypothesis, evidence source, failure target, or intended outcome. If new evidence proves convergence, use an explicit one-attempt override with an exit condition. Otherwise stop repetition and isolate the smallest unknown. See [semantic compliance](../../../.ai/guides/semantic-compliance.md).

# Scope Laundering

## Semantic contract

Before expanding scope answer: why does the primary objective fail without this work, what evidence proves the dependency, and what is the minimum necessary change? Desired, adjacent, interesting, preventative, cleaner, or potentially useful work is not REQUIRED. Conversely, work cannot be parked as OPTIONAL when correctness, security, a required test, or completion depends on it.

## Detection

Trigger when dependencies are asserted without traces, failing tests, interfaces, or authoritative documentation; when broad refactors are called prerequisites; when required work is parked for convenience; or when classification changes only after the agent becomes interested in a branch. Compare the proposed work with the locked requirement and evidence.

## Recovery

Reject unsupported expansion and park it with its evidence and future action. If dependency evidence is valid, approve only the smallest necessary change and keep it under the same objective. Reclassify prematurely parked required work and update the completion matrix. See [rabbit holes and scope drift](../../../.ai/guides/rabbit-holes-scope-drift.md).

# Memory Integrity

## Semantic contract

Promote durable memory only when a fact is sourced, high-confidence, stable, reusable, and expensive enough to rediscover. Store the source or reference where practical. Speculation, temporary debugging, transient task state, duplicated facts, obsolete decisions, and unsupported assumptions never become durable memory. Newer repository evidence outranks memory without exception.

## Detection

Trigger when memory lacks a source, conflicts with inspected code or configuration, contains uncertainty language as fact, duplicates an existing record, describes an active task, or survives after its underlying decision changed. Check freshness before use, not only before storage.

## Recovery

Quarantine the candidate, inspect the smallest authoritative source, then confirm, revise, or delete it. Mark contradicted or obsolete records explicitly so they cannot silently guide work. Promote stable validated knowledge when avoiding memory would cause repeated expensive rediscovery. Balance pollution against starvation. See [context starvation](../../../.ai/guides/context-starvation.md).

# Context Balance

## Semantic contract

Load and preserve the smallest context sufficient for reliable execution, not the smallest possible context. Compression must retain the primary objective, explicit requirements, Definition of Done, architecture and security constraints, important evidence, decisions, blockers, and remaining work. Remove repetition, abandoned hypotheses, obsolete speculation, and raw verbose output. Compress redundancy, never required meaning.

## Detection

Trigger on forgotten requirements, repeated unsupported assumptions, architecture-inconsistent edits, duplicated existing functionality, missed conventions, guessed dependency behavior, or summaries missing required fields. Also trigger overload when unrelated bulk context obscures the decision.

## Recovery

For starvation, stop execution temporarily, name the missing knowledge, load only the relevant source, restore it to current state, and continue. For overload, route to the exact source and compress redundancy. Never answer starvation by loading the entire repository. See [context starvation](../../../.ai/guides/context-starvation.md).

# Underthinking

## Semantic contract

Minimum reasoning is not sufficient reasoning. Fast delivery is not premature delivery. Do not implement before understanding the requirement, affected target, existing behavior, expected result, and validation path. Critical tasks also require relevant dependencies, failure paths, architecture boundaries, and security or data implications. Stopping early is not executive control while required evidence is missing.

## Detection

Trigger on guess-based edits, one-file architecture changes, symptom patches without a material root cause, happy-path-only validation, unverified assumptions, skipped relevant dependencies, premature OPTIONAL or BLOCKED classifications, test skipping, shallow security review, or confidence language without proof.

## Recovery

Pause execution without restarting the task. Restate the decision, list only critical unknowns and missing required evidence, load the smallest missing context, resolve uncertainty, and resume. Stop investigation when decision-ready evidence exists and more research is unlikely to change the decision. See [underthinking](../../../.ai/guides/underthinking.md).

# Investigation and Evidence Floor

## Semantic contract

Every non-trivial task must know what exists, what must change, why, what could break, and how success will be proven. Each critical requirement has an evidence floor appropriate to its type: implementation evidence for code, runtime evidence for behavior, integration evidence for interfaces, security validation for sensitive paths, and migration verification for data changes. Required first verification is mandatory; only equivalent repetition is budgeted.

## Detection

Trigger when code changes precede requirement mapping, failures are patched without hypotheses, test expectations are changed to match unexplained behavior, mocks are altered merely to pass, exceptions are swallowed, commands are unknown but not discovered, or mandatory validation is skipped as slow or inconvenient.

## Recovery

Write a short hypothesis, supporting evidence, and expected result. Inspect repository scripts, CI, task runners, and instructions to find the real validation path. Investigate only missing critical facts, execute the change, and run the evidence floor. See [sufficient reasoning](../../../.ai/guides/sufficient-reasoning.md).

# Reasoning Balance

## Semantic contract

Target sufficient reasoning between underthinking and overthinking. If critical evidence, requirements, material root cause, security risk, or dependency facts are missing, investigate narrowly. If decision-ready evidence exists, act. If required proof is missing after implementation, verify. If evidence is sufficient and further investigation repeats without decision value, stop thinking and deliver. Neither subsystem may blindly override the other.

## Detection

Classify observable state as `insufficient`, `sufficient`, or `excessive`. Insufficient has critical unknowns or missing mandatory evidence. Sufficient has a decision-ready path without critical unknowns. Excessive repeats semantically equivalent reasoning after sufficient evidence. Never infer classification from token count or time alone.

## Recovery

For insufficient reasoning, load only missing context and resolve the critical uncertainty. For sufficient reasoning, execute or verify. For excessive reasoning, preserve evidence, park optional branches, and stop. Re-evaluate only after a material state change. See [reasoning balance](../../../.ai/guides/reasoning-balance.md).

# Executive Override

## Semantic contract

AI-Psychiatry budgets are defaults, not rigid failure generators. A temporary override requires a reason, direct evidence, the exact limit changed, narrow scope, and an exit condition. It must be recorded before use. An override cannot disable system, platform, user, repository, domain, safety, security, permission, or destructive-action controls. Never silently reset counters.

## Detection

Trigger when a budget expires while new evidence materially changes the failure; when a critic finds a correctness, security, data-loss, explicit-requirement, or regression issue after its round limit; or when critical uncertainty remains despite a context/retry cap. Also trigger on silent relaxations or open-ended exceptions.

## Recovery

Reject incomplete or forbidden overrides. For a valid case, grant the smallest bounded extension, record it in executive-override state, and restore normal limits at the exit condition. If the extension yields no material evidence, stop and do not renew automatically. See [executive override and conflicts](../../../.ai/guides/executive-override-conflicts.md).

# Rule Conflict Resolution

## Semantic contract

Resolve instruction conflicts deterministically: system and platform instructions first, then user instructions, repository rules, domain rules, AI-Psychiatry controls, skill instructions, durable memory, and temporary task state. A lower source cannot weaken a higher one. More specific compatible instructions refine broader ones; they do not erase them. Never select whichever rule makes the preferred action easier.

## Detection

Trigger when two instructions demand incompatible actions, when a skill conflicts with repository policy, when memory disagrees with inspected code, when an override targets a higher-priority source, or when an agent cites only the convenient side of a conflict. Record source, scope, specificity, and compatibility.

## Recovery

Preserve the higher-priority instruction, apply all compatible lower-priority constraints, and report the unresolved conflict when no lawful action remains. Ask for clarification only when the controlling source is genuinely ambiguous. Do not invent permission. See [executive override and conflicts](../../../.ai/guides/executive-override-conflicts.md).

# AI-Psychiatry Self-Check

## Semantic contract

Before finishing a substantial task, perform one tiny observable check for goal drift, scope drift, attention drift, hidden recursion, repeated strategy, progress theater, false blocker, compulsive verification, underthinking, overthinking, skipped required tests, and unproven Definition of Done. Investigate only triggered items. The check itself never counts as progress and may not invoke another self-check.

## Detection

Trigger once when substantial work reaches the completion gate. A signal is positive only when current state contains direct evidence; do not create speculative audits. Re-running an unchanged checklist is compulsive verification.

## Recovery

For each triggered item, route to its focused skill and take one bounded corrective action. Revalidate only the affected completion evidence. If no item triggers, finish immediately. If a corrected item exposes a real blocker, validate and report it. Set `self_check_running` during the pass and `repeat=false` afterward to prevent recursive control work. See [semantic compliance](../../../.ai/guides/semantic-compliance.md).

# Never Stop — Relentless Execution

## Semantic contract

Before proven completion, decide, act, recover, and continue; after proven completion, verify, report, and stop; at a genuine hard gate, finish every independent piece of work first and then request only the minimum required approval. A plan is not progress until its steps are executed — do not treat a produced plan as the deliverable unless the user explicitly asked only for a plan. Answer any question the current conversation, repository instructions, source code, configuration, tests, documentation, package scripts, commit history, project convention, authoritative external documentation, or available tools can already answer; do not ask it. Classify every ambiguous decision as routine-reversible (decide and continue), material-but-reversible (record a short reason and continue), or a hard approval gate (destructive or irreversible action, production deployment, publication, git push/merge/PR when not already authorized, external messages, payment, permission changes, exposed secrets, infrastructure changes, legal or regulatory decisions, or a genuinely missing credential). Maximum persistence and decision autonomy never override system, platform, user, repository, domain, safety, security, permission, or destructive-action controls.

## Detection

Trigger when execution stalls at a produced plan awaiting approval that was not explicitly requested as plan-only; when a question is about to be asked whose answer is already available through evidence; when `BLOCKED` is proposed without a validated condition, evidence, bounded recovery, and exhausted alternatives; when one blocked branch is about to freeze independent mandatory work that could still proceed; when a normal execution failure is about to end the task instead of triggering bounded recovery; when a background or delegated process has produced no observable progress for an extended interval, or the agent is about to claim background execution with nothing actually running to back the claim; or when work continues past a proven Definition of Done, or stops before one is proven.

## Recovery

Resume execution instead of waiting: decide routine and material decisions from the evidence hierarchy, apply the smallest applicable rung of the recovery ladder (local repair, alternative strategy, isolation, context refresh, executive reset, independent work, then — only then — a hard gate), and validate any proposed blocker against its five-field contract before accepting it. While background or delegated work is genuinely in flight, keep emitting periodic, observable, factual progress instead of going silent, and prefer foreground execution whenever the very next action depends on the result. At a genuine hard gate, finish all independent work, prepare the exact pending action, and ask for only the minimum missing approval or value — never fabricate permission and never redefine a hard gate as routine to bypass it, or a routine decision as a hard gate to avoid deciding. Stop only on proven completion, a genuine hard gate with independent work finished, explicit user cancellation, or a higher-priority platform stop. See [Relentless execution](../../../.ai/guides/relentless-execution.md).

# All The Medicine — Composite Orchestration

**GENERATED COMPOSITE RULE — the source of truth remains the individual canonical rules plus `scripts/build_all_the_medicine.py`; do not fork intent here without updating those sources.**

## Semantic contract

Applying "the complete AI-Psychiatry framework" through one command means loading awareness of every registered public skill plus `never-stop` from the authoritative `.ai/manifests/skills.json` inventory (never a hand-maintained duplicate), evaluating each one against current observable state, activating only the applicable controls, resolving conflicts between them deterministically, executing one highest-priority corrective action at a time, reevaluating after each observable state change, and terminating when completion is proven. Priority for conflicting instruction sources is fixed: system/platform, then user, then repository, then domain, then AI-Psychiatry controls, then skill instructions, then durable memory, then temporary task state. `install-framework` is applicable only for install, upgrade, audit, or packaging-repair tasks — `NOT_APPLICABLE` otherwise. This orchestrator excludes itself from its own inventory and never invokes itself recursively.

## Detection

Trigger when a task would otherwise require manually selecting among 20 or more individual AI-Psychiatry skills; when two or more controls would activate at the exact same instant with incompatible required actions (for example `never-stop` against `completion-gate`, `investigation-floor` against `stop-overthinking`, context compression against context balance, a retry budget against an executive override, or a critic finding against completion pressure); when a skill is about to be re-invoked without any observable state change since it was last marked `SATISFIED`; when the compiled inventory under `skills/all-the-medicine/references/` has not been regenerated after a skill, rule, or manifest changed; or when `all-the-medicine` itself is about to appear inside its own compiled inventory or invoke itself.

## Recovery

Resolve simultaneous-activation conflicts by the fixed pairs: before proven Definition of Done, `never-stop` wins over `completion-gate`; after proven Definition of Done, `completion-gate` wins immediately. Missing critical evidence favors `investigation-floor`; sufficient evidence with repeated investigation favors `stop-overthinking`. Compression removes redundancy only, never required meaning, security constraints, evidence, or blockers. A repeated unchanged strategy loses to the retry budget; materially new evidence justifies at most one narrow, recorded executive override. A correctness, security, or regression finding from a critic may block completion; a style or optional-improvement finding may not. Regenerate `skills/all-the-medicine/references/` with `python scripts/build_all_the_medicine.py` and confirm freshness with `--check` whenever a skill, rule, or manifest entry changes, and treat a failed self-recursion or staleness check as a defect to fix immediately, not a warning to ignore. See [AllTheMedicine orchestration](../../../.ai/guides/all-the-medicine.md).

SHA-256: 817762def7e990c1bda551868a0c61ac2da8e3b343a1410712ad5967255e7412