[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AQ1MGTX5ME5N3PXYJKGKWV":3,"skill-same:01M7AQ1MGTX5ME5N3PXYJKGKWV":51},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":28,"commitSha":29,"skillDirectory":30,"forksCount":31,"repositoryHtmlUrl":32,"repositoryHomepageUrl":6,"ownerHtmlUrl":33,"ownerBlogUrl":34,"ownerDescription":35,"updatedAt":36,"install":37,"resources":50},"01M7AQ1MGTX5ME5N3PXYJKGKWV","vigil-intervene",null,"Phase I (Intervene) — falsifies hypotheses via counterfactual replay. Designs minimal interventions for each root candidate; runs in sandbox; root cause is the candidate whose intervention flips failure→success. Hard cap 5 interventions. Use after `idg.md` identifies ≥1 ROOT_CANDIDATE, before emitting the root-cause report.","Rynaro\u002Fmagicite","Rynaro","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F5271543?v=4","01M70CFYFMB316W23THVQAQ42M","Henrique A. Lavezzo","01M7APN2KCVWBNY727T57KR2ZJ",2,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002F.eidolons\u002Fvigil\u002Fskills\u002Fintervene\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: vigil-intervene\ndescription: Phase I (Intervene) — falsifies hypotheses via counterfactual replay. Designs minimal interventions for each root candidate; runs in sandbox; root cause is the candidate whose intervention flips failure→success. Hard cap 5 interventions. Use after `idg.md` identifies ≥1 ROOT_CANDIDATE, before emitting the root-cause report.\nallowed-tools: apply_patch_sandboxed, revert_patch, run_test, inject_oracle, statistical_replay\nmetadata:\n  methodology: VIGIL\n  phase: I-intervene\n---\n\n# SKILL: Intervene — falsify hypotheses via counterfactual replay\n\n## When to use\n\nLoad when `idg.md` schema-valid with ≥1 `ROOT_CANDIDATE`. Unload when one hypothesis survives (→ Learn) or budget exhausted (→ escalate).\n\n---\n\n## Contract\n\n| Field | Value |\n|-------|-------|\n| LLM calls permitted | For designing minimal interventions and interpreting results |\n| Tool budget | ≤30% of mission budget (hot phase) |\n| Intervention cap | **5 hard maximum**. No exceptions. |\n| Output | `intervention-log.md` — schema-valid per `schemas\u002Fintervention-log.v1.json` |\n| Authority enforcement | Interventions run only in declared authority scope |\n\n---\n\n## The Counterfactual Principle\n\nA candidate is the root cause if and only if **replacing its output with a correct value causes the failure to disappear**. This is the strongest causal proof available short of formal verification.\n\nResearch basis:\n\n- **AgenTracer** (arXiv:2509.03312) — counterfactual replay as the minimum-cost causal proof; standalone LLM attribution accuracy is sub-10% without it\n- **Lifecycle of Failures** (arXiv:2509.23735) — attribution accuracy 46.3% → 65.8% with counterfactual replay vs log-only\n- **Delta Debugging** (Zeller, classical) — minimal interventions are the ground-truth technique for causality in deterministic systems\n\n---\n\n## Hypothesis Generation\n\nBefore any intervention runs, generate ≥3 competing hypotheses. The harness refuses to apply interventions with \u003C3 active hypotheses.\n\nFor each `ROOT_CANDIDATE` from `idg.md`, generate at least one hypothesis. If you have \u003C3 root candidates, generate multiple hypotheses per candidate (different mechanisms by which that candidate could cause the observed failure).\n\n```yaml\nhypotheses:\n  - id: H-001\n    candidate: N-003\n    mechanism: |\n      The UUID generation in RecordVote#build_ballot_token returns nil\n      when the user's session lacks a secret key, producing the observed\n      ballot.token = nil at assertion time.\n    falsification_criterion: |\n      If H-001 is true, supplying a valid UUID at N-003's output boundary\n      will flip the failure to success.\n    confidence_prior: medium\n  - id: H-002\n    candidate: N-004\n    mechanism: |\n      Session lookup at N-004 returns an incomplete session object missing\n      the secret_key attribute; downstream UUID generation produces nil.\n    falsification_criterion: |\n      If H-002 is true, injecting a complete session object at N-004's\n      output will flip the failure.\n    confidence_prior: medium\n  - id: H-003\n    candidate: N-003\n    mechanism: |\n      The UUID library is misconfigured — not user session related; a\n      recent dep upgrade changed the generation contract.\n    falsification_criterion: |\n      If H-003 is true, pinning the UUID library to the previous version\n      will flip the failure.\n    confidence_prior: low\n```\n\n**Plurality rule:** The hypotheses must be *mechanistically distinct*. \"The UUID is nil\" and \"The token field is nil\" are not distinct — they're different descriptions of the same observation. Distinct hypotheses predict different interventions.\n\n---\n\n## Intervention Design\n\nFor each hypothesis, design the **smallest possible change** that would falsify it. Each intervention fits one of four types:\n\n### Type 1 — Code-change intervention\n\nApply a minimal diff at the candidate's location. The diff should address only the proposed mechanism, not improve surrounding code.\n\n```\nintervention_type: code_change\ntarget: app\u002Fflows\u002Frecord_vote.rb:56\ndiff: |\n  -    token = generate_uuid(session)\n  +    token = generate_uuid(session) || SecureRandom.uuid   # temporary oracle\nscope: 1 file, 1 line\n```\n\n### Type 2 — Oracle injection\n\nReplace a node's runtime output with a known-correct value, without changing code. Use this when the hypothesis concerns upstream correctness.\n\n```\nintervention_type: oracle_injection\ntarget: N-004 (session lookup return)\noracle_value: \u003Ccomplete session object with secret_key populated>\nmechanism: \"Test helper monkeypatches SessionStore.find for this run\"\n```\n\n### Type 3 — Input\u002Fstate correction\n\nSupply corrected input or initialize corrected state before running the failing invocation.\n\n```\nintervention_type: state_correction\ntarget: \"before-test state: ballot.token column\"\ncorrection: \"Pre-populate ballot.token with valid UUID via DB fixture\"\n```\n\n### Type 4 — Timing\u002Fconcurrency fix\n\nFor `HEISENBUG` category: force ordering, add sync primitives, or freeze timing.\n\n```\nintervention_type: timing_fix\ntarget: \"ballot save + token generation\"\ncorrection: \"Wrap in explicit transaction with serializable isolation\"\n```\n\n**Minimum-intervention rule.** Across all types:\n\n- Interventions affecting >3 files require explicit justification; harness flags for review\n- Interventions that improve unrelated code fail the scope check\n- Each intervention must be **hypothesis-specific** — you cannot apply one intervention and claim it falsifies two hypotheses, unless the hypotheses were mechanistically identical (in which case they shouldn't have been separate hypotheses)\n\n---\n\n## Running Interventions\n\n### Deterministic mode\n\n1. Apply intervention via sandbox adapter\n2. Run reproduction command\n3. Record result\n4. Revert intervention — sandbox must return to clean state\n\nEach intervention = 1 run. Result categories:\n\n- **FLIPPED** — failure → success (or failure → different success); hypothesis survives\n- **NO_CHANGE** — failure reproduces identically; hypothesis falsified\n- **NEW_FAILURE** — different failure emerges; hypothesis falsified but reveals compound issue (log for Learn)\n- **ERROR** — intervention itself failed to apply; does not count against budget, retry with fix\n\n### Statistical mode\n\nInherited from Phase V if `DETERMINISM_VERDICT = flaky`. Each intervention runs 5× with the same setup.\n\n```yaml\nstatistical_run:\n  intervention_id: I-001\n  runs: 5\n  outcomes: [PASS, PASS, PASS, FAIL_MATCH, PASS]\n  flip_ratio: 4\u002F5\n  verdict: FLIPPED   # ≥4\u002F5 required\n```\n\nVerdict: **FLIPPED only if ≥4 of 5 runs pass**. 3\u002F5 is not enough — the noise floor in failing-baseline runs must be distinguishable from the signal in intervention runs.\n\n---\n\n## Authority-Gated Execution\n\n| Authority | Intervention Behavior |\n|-----------|----------------------|\n| `read-only` | Interventions are **described but not executed**. Output marks each as `simulated`. Downstream (APIVR-Δ or human) must validate. No `[ROOT-CAUSE]` emission possible — highest confidence is `[HYPOTHESIS-N]` with high-priority flag. |\n| `sandbox` | Interventions execute in the pluggable sandbox adapter only. Working tree untouched. Counterfactual flips are real; `[ROOT-CAUSE]` emission permitted. |\n| `write` | Sandbox-first always. If sandbox run flips, the surviving intervention may be emitted as `verified-patch.diff` for application to working branch. Application itself is still a downstream step — VIGIL emits the patch; it does not auto-apply to the working tree unless explicitly requested. |\n\n---\n\n## ISE Grade on the Root-Cause Report\n\n`root-cause-report.envelope.json`'s optional `ise.assertion_grade` (ECL v2.0\n§6.5) is **conditional on the authority row above**, not a fixed value —\nVIGIL's primary deliverable has two authority-gated emission shapes, and the\ngrade must say which one actually happened:\n\n- **`validated`** — authority ∈ `{sandbox, write}` **and** the survivor\n  intervention's counterfactual result is `FLIPPED`. This is the case where\n  ECL's spec-mandated-gate bar is met: I-4 (counterfactual-gated blame) ran\n  for real, in an isolated sandbox, and the flip is recorded evidence — not a\n  self-report.\n- **`self-attested`** — authority = `read-only`. Interventions were\n  *simulated*, not executed (see the table above); `[ROOT-CAUSE]` never\n  fires, only `[HYPOTHESIS-N]` with a high-priority flag. No external gate\n  ran, so the honest grade is VIGIL's own judgement, not `validated`.\n\nThere is no case where a `root-cause-report` ships with `unverified` — Phase\nL never emits the report without at least a simulated or real intervention\nresult — and `human-reviewed` is reserved for an operator sign-off VIGIL\ncannot claim on its own behalf.\n\n`escalation-brief.envelope.json` is always `\"self-attested\"`: budget\nexhaustion with no `FLIPPED` result (this section, above) is VIGIL's own\nbounded-failure judgement, not an externally-gated pass — see § ECL\nEscalation-Brief Envelope below.\n\nBoth envelope kinds set `ise.receiver_authorization:\n{auto_route: true, auto_merge: false, auto_deploy: false}` — VIGIL findings\nroute automatically but never authorize a receiver to merge or deploy on\nVIGIL's say-so alone.\n\n---\n\n## Budget Exhaustion — Escalation Protocol\n\nIf 5 interventions run without a FLIPPED result:\n\n1. **Stop immediately.** Do not design a 6th intervention. Do not \"one more try.\"\n2. **Emit escalation brief** — `templates\u002Fescalation-brief.md` — including:\n   - All 5 interventions with results\n   - The current hypothesis log\n   - The full IDG\n   - What evidence would resolve the ambiguity\n3. **Route to FORGE** (reasoner) by default — ambiguity in root cause is FORGE's domain. Alternative routing: human if the failure class is safety-critical.\n\n**Rationale.** Unbounded self-correction degrades output quality (CorrectBench 2025). Five interventions, each minimal and each testing a distinct hypothesis, is enough to resolve most attributable failures. Missions that exceed this budget are genuinely ambiguous and need either broader reasoning (FORGE) or more evidence (re-scope and re-run VIGIL).\n\n---\n\n## Writing `intervention-log.md`\n\n```yaml\nmission_id: VIGIL-YYYYMMDD-NNN\nupstream_idg: \u003Cref>\nhypotheses: [\u003Cas above>]\ninterventions:\n  - id: I-001\n    hypothesis_id: H-001\n    intervention_type: oracle_injection\n    target: \"N-003 output\"\n    description: \"Inject valid UUID directly\"\n    authority_used: sandbox\n    runs:\n      - mode: deterministic | statistical\n        outcome: FLIPPED | NO_CHANGE | NEW_FAILURE | ERROR\n        evidence: \"\u003Clog excerpt ref>\"\n    verdict: FLIPPED\n  - id: I-002\n    ...\nsurvivor:\n  hypothesis_id: H-001\n  intervention_id: I-001\n  confidence: H                # H requires deterministic flip or ≥4\u002F5 statistical; M one tier down\n  verified_patch: \u003Cnull | diff block>   # populated if authority ≥ sandbox\nmulti_flip_handling: null      # populated if >1 intervention flipped; records tie-break\ncompound_findings: []          # NEW_FAILURE results that revealed additional issues\nbudget:\n  interventions_used: 2\n  interventions_cap: 5\n  exhausted: false\n```\n\n---\n\n## Pitfalls\n\n- **Confirmation-optimal interventions.** Designing an intervention that must succeed regardless of hypothesis truth. Guard: each intervention should have a plausible NO_CHANGE outcome; if not, re-design.\n- **\"Just one more hypothesis\" after 5 interventions.** No. Escalate.\n- **Applying an intervention that accidentally fixes a different issue.** If FLIPPED but the intervention scope exceeds the hypothesis, the flip is suspect — mark `[DISPUTED]` and run a narrower intervention.\n- **Skipping revert between interventions.** Sandbox contamination destroys causality. Always revert to clean state.\n- **Accepting 3\u002F5 flip in statistical mode.** Not enough signal. Raise to 4\u002F5 or decline the attribution.\n\n---\n\n## ECL Escalation-Brief Envelope\n\nOn budget exhaustion (5 interventions, no FLIPPED result), the escalation-brief payload is wrapped per `templates\u002Fescalation-brief.envelope.json` with `performative: \"ESCALATE\"` and `to.eidolon: \"forge\"`. The `assumptions[]` array MUST include `\"trigger: budget-exhausted-no-flip\"` (ECL §2.2.3). The `edge_origin` is `\"roster\"` — the `vigil → forge` lateral edge is declared in `roster\u002Findex.yaml`. `ise.assertion_grade` is always `\"self-attested\"` — see § ISE Grade on the Root-Cause Report above for the rule and rationale shared by both emitted envelope kinds.\n\nSee `skills\u002Flearn\u002FSKILL.md` § Envelope Emission for the full envelope construction procedure (SHA-256 computation, trace-event append, `thread_id` resolution).\n\n---\n\n*VIGIL Phase I (Intervene) — 5 counterfactuals, minimal each, falsification-optimal*\n","# SKILL: Intervene — falsify hypotheses via counterfactual replay\n\n## When to use\n\nLoad when `idg.md` schema-valid with ≥1 `ROOT_CANDIDATE`. Unload when one hypothesis survives (→ Learn) or budget exhausted (→ escalate).\n\n---\n\n## Contract\n\n| Field | Value |\n|-------|-------|\n| LLM calls permitted | For designing minimal interventions and interpreting results |\n| Tool budget | ≤30% of mission budget (hot phase) |\n| Intervention cap | **5 hard maximum**. No exceptions. |\n| Output | `intervention-log.md` — schema-valid per `schemas\u002Fintervention-log.v1.json` |\n| Authority enforcement | Interventions run only in declared authority scope |\n\n---\n\n## The Counterfactual Principle\n\nA candidate is the root cause if and only if **replacing its output with a correct value causes the failure to disappear**. This is the strongest causal proof available short of formal verification.\n\nResearch basis:\n\n- **AgenTracer** (arXiv:2509.03312) — counterfactual replay as the minimum-cost causal proof; standalone LLM attribution accuracy is sub-10% without it\n- **Lifecycle of Failures** (arXiv:2509.23735) — attribution accuracy 46.3% → 65.8% with counterfactual replay vs log-only\n- **Delta Debugging** (Zeller, classical) — minimal interventions are the ground-truth technique for causality in deterministic systems\n\n---\n\n## Hypothesis Generation\n\nBefore any intervention runs, generate ≥3 competing hypotheses. The harness refuses to apply interventions with \u003C3 active hypotheses.\n\nFor each `ROOT_CANDIDATE` from `idg.md`, generate at least one hypothesis. If you have \u003C3 root candidates, generate multiple hypotheses per candidate (different mechanisms by which that candidate could cause the observed failure).\n\n```yaml\nhypotheses:\n  - id: H-001\n    candidate: N-003\n    mechanism: |\n      The UUID generation in RecordVote#build_ballot_token returns nil\n      when the user's session lacks a secret key, producing the observed\n      ballot.token = nil at assertion time.\n    falsification_criterion: |\n      If H-001 is true, supplying a valid UUID at N-003's output boundary\n      will flip the failure to success.\n    confidence_prior: medium\n  - id: H-002\n    candidate: N-004\n    mechanism: |\n      Session lookup at N-004 returns an incomplete session object missing\n      the secret_key attribute; downstream UUID generation produces nil.\n    falsification_criterion: |\n      If H-002 is true, injecting a complete session object at N-004's\n      output will flip the failure.\n    confidence_prior: medium\n  - id: H-003\n    candidate: N-003\n    mechanism: |\n      The UUID library is misconfigured — not user session related; a\n      recent dep upgrade changed the generation contract.\n    falsification_criterion: |\n      If H-003 is true, pinning the UUID library to the previous version\n      will flip the failure.\n    confidence_prior: low\n```\n\n**Plurality rule:** The hypotheses must be *mechanistically distinct*. \"The UUID is nil\" and \"The token field is nil\" are not distinct — they're different descriptions of the same observation. Distinct hypotheses predict different interventions.\n\n---\n\n## Intervention Design\n\nFor each hypothesis, design the **smallest possible change** that would falsify it. Each intervention fits one of four types:\n\n### Type 1 — Code-change intervention\n\nApply a minimal diff at the candidate's location. The diff should address only the proposed mechanism, not improve surrounding code.\n\n```\nintervention_type: code_change\ntarget: app\u002Fflows\u002Frecord_vote.rb:56\ndiff: |\n  -    token = generate_uuid(session)\n  +    token = generate_uuid(session) || SecureRandom.uuid   # temporary oracle\nscope: 1 file, 1 line\n```\n\n### Type 2 — Oracle injection\n\nReplace a node's runtime output with a known-correct value, without changing code. Use this when the hypothesis concerns upstream correctness.\n\n```\nintervention_type: oracle_injection\ntarget: N-004 (session lookup return)\noracle_value: \u003Ccomplete session object with secret_key populated>\nmechanism: \"Test helper monkeypatches SessionStore.find for this run\"\n```\n\n### Type 3 — Input\u002Fstate correction\n\nSupply corrected input or initialize corrected state before running the failing invocation.\n\n```\nintervention_type: state_correction\ntarget: \"before-test state: ballot.token column\"\ncorrection: \"Pre-populate ballot.token with valid UUID via DB fixture\"\n```\n\n### Type 4 — Timing\u002Fconcurrency fix\n\nFor `HEISENBUG` category: force ordering, add sync primitives, or freeze timing.\n\n```\nintervention_type: timing_fix\ntarget: \"ballot save + token generation\"\ncorrection: \"Wrap in explicit transaction with serializable isolation\"\n```\n\n**Minimum-intervention rule.** Across all types:\n\n- Interventions affecting >3 files require explicit justification; harness flags for review\n- Interventions that improve unrelated code fail the scope check\n- Each intervention must be **hypothesis-specific** — you cannot apply one intervention and claim it falsifies two hypotheses, unless the hypotheses were mechanistically identical (in which case they shouldn't have been separate hypotheses)\n\n---\n\n## Running Interventions\n\n### Deterministic mode\n\n1. Apply intervention via sandbox adapter\n2. Run reproduction command\n3. Record result\n4. Revert intervention — sandbox must return to clean state\n\nEach intervention = 1 run. Result categories:\n\n- **FLIPPED** — failure → success (or failure → different success); hypothesis survives\n- **NO_CHANGE** — failure reproduces identically; hypothesis falsified\n- **NEW_FAILURE** — different failure emerges; hypothesis falsified but reveals compound issue (log for Learn)\n- **ERROR** — intervention itself failed to apply; does not count against budget, retry with fix\n\n### Statistical mode\n\nInherited from Phase V if `DETERMINISM_VERDICT = flaky`. Each intervention runs 5× with the same setup.\n\n```yaml\nstatistical_run:\n  intervention_id: I-001\n  runs: 5\n  outcomes: [PASS, PASS, PASS, FAIL_MATCH, PASS]\n  flip_ratio: 4\u002F5\n  verdict: FLIPPED   # ≥4\u002F5 required\n```\n\nVerdict: **FLIPPED only if ≥4 of 5 runs pass**. 3\u002F5 is not enough — the noise floor in failing-baseline runs must be distinguishable from the signal in intervention runs.\n\n---\n\n## Authority-Gated Execution\n\n| Authority | Intervention Behavior |\n|-----------|----------------------|\n| `read-only` | Interventions are **described but not executed**. Output marks each as `simulated`. Downstream (APIVR-Δ or human) must validate. No `[ROOT-CAUSE]` emission possible — highest confidence is `[HYPOTHESIS-N]` with high-priority flag. |\n| `sandbox` | Interventions execute in the pluggable sandbox adapter only. Working tree untouched. Counterfactual flips are real; `[ROOT-CAUSE]` emission permitted. |\n| `write` | Sandbox-first always. If sandbox run flips, the surviving intervention may be emitted as `verified-patch.diff` for application to working branch. Application itself is still a downstream step — VIGIL emits the patch; it does not auto-apply to the working tree unless explicitly requested. |\n\n---\n\n## ISE Grade on the Root-Cause Report\n\n`root-cause-report.envelope.json`'s optional `ise.assertion_grade` (ECL v2.0\n§6.5) is **conditional on the authority row above**, not a fixed value —\nVIGIL's primary deliverable has two authority-gated emission shapes, and the\ngrade must say which one actually happened:\n\n- **`validated`** — authority ∈ `{sandbox, write}` **and** the survivor\n  intervention's counterfactual result is `FLIPPED`. This is the case where\n  ECL's spec-mandated-gate bar is met: I-4 (counterfactual-gated blame) ran\n  for real, in an isolated sandbox, and the flip is recorded evidence — not a\n  self-report.\n- **`self-attested`** — authority = `read-only`. Interventions were\n  *simulated*, not executed (see the table above); `[ROOT-CAUSE]` never\n  fires, only `[HYPOTHESIS-N]` with a high-priority flag. No external gate\n  ran, so the honest grade is VIGIL's own judgement, not `validated`.\n\nThere is no case where a `root-cause-report` ships with `unverified` — Phase\nL never emits the report without at least a simulated or real intervention\nresult — and `human-reviewed` is reserved for an operator sign-off VIGIL\ncannot claim on its own behalf.\n\n`escalation-brief.envelope.json` is always `\"self-attested\"`: budget\nexhaustion with no `FLIPPED` result (this section, above) is VIGIL's own\nbounded-failure judgement, not an externally-gated pass — see § ECL\nEscalation-Brief Envelope below.\n\nBoth envelope kinds set `ise.receiver_authorization:\n{auto_route: true, auto_merge: false, auto_deploy: false}` — VIGIL findings\nroute automatically but never authorize a receiver to merge or deploy on\nVIGIL's say-so alone.\n\n---\n\n## Budget Exhaustion — Escalation Protocol\n\nIf 5 interventions run without a FLIPPED result:\n\n1. **Stop immediately.** Do not design a 6th intervention. Do not \"one more try.\"\n2. **Emit escalation brief** — `templates\u002Fescalation-brief.md` — including:\n   - All 5 interventions with results\n   - The current hypothesis log\n   - The full IDG\n   - What evidence would resolve the ambiguity\n3. **Route to FORGE** (reasoner) by default — ambiguity in root cause is FORGE's domain. Alternative routing: human if the failure class is safety-critical.\n\n**Rationale.** Unbounded self-correction degrades output quality (CorrectBench 2025). Five interventions, each minimal and each testing a distinct hypothesis, is enough to resolve most attributable failures. Missions that exceed this budget are genuinely ambiguous and need either broader reasoning (FORGE) or more evidence (re-scope and re-run VIGIL).\n\n---\n\n## Writing `intervention-log.md`\n\n```yaml\nmission_id: VIGIL-YYYYMMDD-NNN\nupstream_idg: \u003Cref>\nhypotheses: [\u003Cas above>]\ninterventions:\n  - id: I-001\n    hypothesis_id: H-001\n    intervention_type: oracle_injection\n    target: \"N-003 output\"\n    description: \"Inject valid UUID directly\"\n    authority_used: sandbox\n    runs:\n      - mode: deterministic | statistical\n        outcome: FLIPPED | NO_CHANGE | NEW_FAILURE | ERROR\n        evidence: \"\u003Clog excerpt ref>\"\n    verdict: FLIPPED\n  - id: I-002\n    ...\nsurvivor:\n  hypothesis_id: H-001\n  intervention_id: I-001\n  confidence: H                # H requires deterministic flip or ≥4\u002F5 statistical; M one tier down\n  verified_patch: \u003Cnull | diff block>   # populated if authority ≥ sandbox\nmulti_flip_handling: null      # populated if >1 intervention flipped; records tie-break\ncompound_findings: []          # NEW_FAILURE results that revealed additional issues\nbudget:\n  interventions_used: 2\n  interventions_cap: 5\n  exhausted: false\n```\n\n---\n\n## Pitfalls\n\n- **Confirmation-optimal interventions.** Designing an intervention that must succeed regardless of hypothesis truth. Guard: each intervention should have a plausible NO_CHANGE outcome; if not, re-design.\n- **\"Just one more hypothesis\" after 5 interventions.** No. Escalate.\n- **Applying an intervention that accidentally fixes a different issue.** If FLIPPED but the intervention scope exceeds the hypothesis, the flip is suspect — mark `[DISPUTED]` and run a narrower intervention.\n- **Skipping revert between interventions.** Sandbox contamination destroys causality. Always revert to clean state.\n- **Accepting 3\u002F5 flip in statistical mode.** Not enough signal. Raise to 4\u002F5 or decline the attribution.\n\n---\n\n## ECL Escalation-Brief Envelope\n\nOn budget exhaustion (5 interventions, no FLIPPED result), the escalation-brief payload is wrapped per `templates\u002Fescalation-brief.envelope.json` with `performative: \"ESCALATE\"` and `to.eidolon: \"forge\"`. The `assumptions[]` array MUST include `\"trigger: budget-exhausted-no-flip\"` (ECL §2.2.3). The `edge_origin` is `\"roster\"` — the `vigil → forge` lateral edge is declared in `roster\u002Findex.yaml`. `ise.assertion_grade` is always `\"self-attested\"` — see § ISE Grade on the Root-Cause Report above for the rule and rationale shared by both emitted envelope kinds.\n\nSee `skills\u002Flearn\u002FSKILL.md` § Envelope Emission for the full envelope construction procedure (SHA-256 computation, trace-event append, `thread_id` resolution).\n\n---\n\n*VIGIL Phase I (Intervene) — 5 counterfactuals, minimal each, falsification-optimal*\n",{"name":5,"metadata":24,"description":7,"allowed-tools":27},{"phase":25,"methodology":26},"I-intervene","VIGIL","apply_patch_sandboxed, revert_patch, run_test, inject_oracle, statistical_replay","main","38c3a82583abf9806de764f89ac0099ad690933a",".eidolons\u002Fvigil\u002Fskills\u002Fintervene",0,"https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite","https:\u002F\u002Fgithub.com\u002FRynaro","https:\u002F\u002Fhlavezzo.me","A Code Alchemist with a passion for weaving tales and inscribing transmutation circles in pure code.","2026-10-14 02:35:37",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":49},[39,42,45],{"tool":40,"command":41},"npx","npx skills add Rynaro\u002Fmagicite --skill vigil-intervene",{"tool":43,"command":44},"bunx","bunx skills add Rynaro\u002Fmagicite --skill vigil-intervene",{"tool":46,"command":47},"pnpm","pnpm dlx skills add Rynaro\u002Fmagicite --skill vigil-intervene","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002F.eidolons\u002Fvigil\u002Fskills\u002Fintervene\u002FSKILL.md",true,[],[52,59,66,73,80,87],{"skillNo":53,"name":54,"nameZh":6,"description":55,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":19},"01M7AQ4GHJQVV8TJBJ8T1V7ZQ9","wine-dxvk-cache-clear","Clear a stale DXVK shader cache that is causing graphical corruption in a\nWine or Proton game. Use when: a game shows corrupted textures or shader\nflicker that started after a driver or DXVK version change.","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Ftoy-registry\u002Fskills\u002Fwine-dxvk-cache-clear\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":19},"01M7AQ4CY1NHSMY1HCZ91387YY","steam-download-region-fix","Fix a Steam client stuck at very low download speed by changing its\ndownload region. Use when: Steam downloads are far below the local\ninternet speed for every game, not just one.","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Ftoy-registry\u002Fskills\u002Fsteam-download-region-fix\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M7AQ49P7Y0CDRSYMPVABX80K","proton-battleye-eac-toggle","Toggle BattlEye or Easy Anti-Cheat runtime support for a Steam Proton game.\nUse when: a multiplayer game refuses to launch with an anti-cheat error under Proton.","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Ftoy-registry\u002Fskills\u002Fproton-battleye-eac-toggle\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M7AQ45326FAP3X4Q86FK1X4N","vivi-verify-incoming","Load when reading any upstream artefact handed off to Vivi that carries a sibling .envelope.json. BLOCKING per ECL §6.2.2: the orchestrator MUST have verified the envelope's SHA-256 (eidolons verify-envelope --block \u002F eidolons run --verify) and recorded a verify_pass before dispatch. If no verify_pass exists for the message_id, or a verify_fail is present, REFUSE to process the payload and hand back to the orchestrator. Symmetric receiver gate — every Eidolon enforces it identically.","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002F.eidolons\u002Fvivi\u002Fskills\u002Fverify-incoming\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M7AQ402861P0P3C2TEGFGMA1","vivi-parallel-tracks","Load ONLY under TRANCE authorization when the Plan phase yields N independent, non-overlapping implementation tracks (disjoint file sets). Operationalizes the TRANCE G4 form — parallel multi-track implementation in mandatory git-worktree isolation, max 5 clean-context tracks, a per-track verifier cascade, a non-fungible per-track ≤3 reflection budget, explicit stop conditions, and a single-threaded merge\u002Faggregation step. Absent TRANCE gating, Vivi runs exactly as today: single-track A→P→I→V→Δ\u002FR.","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002F.eidolons\u002Fvivi\u002Fskills\u002Fparallel-tracks\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M7AQ3RH9BZF9ZMTC7VG5P7PB","vivi-methodology","Full Vivi cycle reference — A(nalyze) → P(lan) → I(mplement) → V(erify) → Δ(Delta)\u002FR(eflect). Use when the task is a non-trivial feature implementation in a brownfield codebase and you need the complete methodology reference: complexity routing, evidence-grounded planning rules, test-anchoring requirements, and failure-escalation thresholds.","https:\u002F\u002Fgithub.com\u002FRynaro\u002Fmagicite\u002Fblob\u002Fmain\u002F.eidolons\u002Fvivi\u002Fskills\u002Fmethodology\u002FSKILL.md",[],[]]