Goal-Recovery Skill | Agent Skills

Goal-Recovery

Thin intent-recovery step. goal-recovery does not re-implement objectives-synthesis, session state-reading, or seed-schema — it composes postflight's P2-DEBRIEF synthesis and the continuation-seed contract, and contributes only what the family lacks: recovery of an unstated goal from execution state, made uncertainty-aware and emitted as a typed handoff-as-prompt envelope. The novel step the loop/goal family was missing.

Provenance / lineage: the genuine ~30% residual of the operator's recurring goal-loop contract ooda --goal==$(detecte/resgate [motivacoes, DoR, contexto, escopo, objetivos] --type handoff-as-prompt). The family already EMITS goals (postflight P3), ANCHORS them to tickets (preflight R0), and CONSUMES them (gap-loop/quiesce) — none RECOVERED an unstated one at session start. Named via anima (agent-register): goal-recovery (rejected intent-recovery — drifts from the family's {{goal}} anchor + the operator's literal --goal; rejected anamnesis — soul-name obscurity in an agent slug). Authored by Claude Opus 4.8 (1M) under operator /enhance directive 2026-07-12; reviewed by @ekson73.

§0 — BEING > Rules (foundational)

Serves the operator's intent. If a phase/gate obstructs recovering the right goal NOW, skip it, log Skipped <phase> — BEING > Rules, proceed. HUMAN_DOMAIN (secrets · production PII · irreversibles · cross-org) recovered as constraints go into guardrails, never actioned here. Never fabricate an intent to look decisive — an ungrounded recovery is inconclusive->HITL, not a confident guess (anti-theater R3/R4). This is the single hard discipline.

When to use

A session's goal is unstated / implicit and a driven loop (ooda-loop / gap-loop / quiesce) needs a typed {{goal}}.
Cold re-entry (amnesic agent, post-/compact, resumed session) — recover what this session is doing before acting.
Any tool that needs the operator's handoff-as-prompt bundle (motivations · DoR · scope · objective tree).

When not to use

The operator stated the goal explicitly this turn — use it directly (S2); recovery is for the UNSTATED.
Emitting a seed at session END (that is postflight P3 — the forward direction).
Anchoring to a known ticket only (that is preflight R0).
Single-shot Q&A / trivial edit — no goal to recover.

Trigger Phrases

"goal-recovery" / "recover the goal" / "resgate o goal" / "detecte o objetivo da sessao"
"what is this session actually trying to do" / "recover intent" / "handoff-as-prompt"

The inference ladder (Skopos recon-first — strongest evidence first)

Probe each source's PRESENCE before inferring (recovered_from[].read records existed-or-absent — auditable). Recover from the strongest available; corroborate across sources; a single source never over-rides an explicit conflict.

1. postflight_seed     a prior session.continuation seed for this branch/session   (the exact inverse artifact)
2. braindump           a raw operator dump supplied AS the invocation subject       (stated-but-unstructured
                       (path or inline text)                                         intent; supersedes weaker
                                                                                     inferred sources for the
                                                                                     work it describes)
3. ticket              preflight R0 anchor — Jira/Linear/GH issue title+body+DoD    (stated intent, if any)
4. objectives_ntree    postflight P2 objectives-map / work-compass CPT N-Tree        (synthesized objectives)
5. ash_journal         agentic-session-harness journal: top-level `goal`, decisions[], spec_alignment drift
6. git_branch          branch name (feat/<slug>) + PR title                          (weak but cheap signal)
7. git_last_commit     last commit subject/body                                      (weak signal)
8. transcript          this session's transcript (transcript_hash)                   (last resort; expensive)

braindump is NOT session residue like the other sources — it is an artifact the operator hands in. It is ranked high because it is the operator's own words about this work, but it is still recovery, not restatement: the goal is present yet unstated-as-such, buried under mixed directives and session-meta. Drop the session-meta (/enhance wrappers, pep-talk, cartesian resource lists offered as examples) before synthesizing — a listed resource is not a requirement to use it. Primary consumer: skills/refine-braindump-to-prompt PHASE 1.

Pipeline (0 -> 4)

0 · DoR       is there a session/context to recover from? (any ladder source present). If NONE and no
              operator prompt -> inconclusive{flag:true, reasons:[no_context_anchor]} -> HITL. Abort-with-reason.
1 · RECON     probe the ladder (Skopos): record which sources exist -> recovered_from[]. Read the strongest
              present source(s). NO fabrication — only what the sources say.
2 · SYNTHESIZE compose postflight P2-DEBRIEF objectives-synthesis over the read sources -> the intent bundle:
              motivations · DoR · context · scope{in,out} · objectives{originating,primary,secondary,auxiliary}.
              Distinguish `originating` (why we started at all) from `primary` (the top deliverable).
3 · HYPOTHESIZE enumerate RANKED goal-interpretations (uncertainty-aware). Each carries confidence + the
              evidence that supports it. The chosen `goal` = hypotheses[0].goal. Aggregate `confidence`.
              confidence < conf_inconclusive (0.60) OR two top hypotheses within delta -> inconclusive.
4 · EMIT      write the `handoff-as-prompt` envelope (templates/handoff-as-prompt.schema.json). VALIDATE it
              (bin/validate_envelope.py). inconclusive.flag=true -> print the ranked hypotheses + STOP-HITL.
              Otherwise emit the validated envelope for the caller (ooda-loop/gap-loop/quiesce/preflight/...).

Uncertainty-aware discipline (the defensive core — do NOT skip)

Ranked hypotheses, not a frozen string. The output carries alternatives + confidence (GOOD arXiv:2508.15119; RECAP 2509.04472). A downstream loop MAY revise goal to a lower-ranked hypothesis mid-run rather than freeze it (lost-in-multi-turn defense).
inconclusive -> HITL. Below-threshold confidence, conflicting top hypotheses, no context anchor, ambiguous scope, or stale state => inconclusive.flag=true; escalate WITH the ranked hypotheses attached (not a blank question). Never drive a loop on a low-confidence goal.
Grounded or absent. Every recovered claim traces to a recovered_from source (anti-theater grounding). A confident recovery (confidence>0) with empty recovered_from is a SpecError.

Hybrid boundary (deterministic envelope / probabilistic content)

| Layer | Determinism | Mechanism | |---|---|---| | intent EXTRACTION | probabilistic | LLM over the inference ladder, uncertainty-aware | | the ENVELOPE | deterministic | handoff-as-prompt.schema.json + validate_envelope.py (typed, range-checked, gate-checked) | | IDEMPOTENCY | deterministic | same session-state -> same validated envelope, modulo the confidence float (banded in the --digest canonicalization) |

Composition (the wiring — DRY; every step lands on an existing primitive)

| Step | Composes (existing — reimplements nothing) | |---|---| | RECON | skills/pulse state-refresh · skills/preflight R0 ticket-anchor · skills/work-compass CPT N-Tree · git/jq reads | | SYNTHESIZE | skills/postflight P2-DEBRIEF objectives-synthesis · continuation-seed contract (the schema base) | | audit source | skills/agentic-session-harness journal (goal, decisions[], spec_alignment) | | EMIT + validate | templates/handoff-as-prompt.schema.json (extends continuation-seed) + bin/validate_envelope.py |

Consumers (independently reusable — Strata elevate)

ooda-loop (its Observe step) · preflight (recover intent when no ticket anchor) · postflight (cross-check emitted seed vs recovered intent — drift signal) · morning-briefing · session-fission (per-node intent). One recovery engine, many callers.

Override parameters

| Flag | Default | Allowed / Notes | |---|---|---| | "<hint>" (positional) | empty | optional operator hint appended to the recovery (does NOT override the ladder evidence). Exception — the braindump source (rank 2): when a braindump path or text is supplied here, the positional IS the rank-2 evidence, not a hint, and it outranks every source below it. A braindump invocation with an empty positional has no rank-2 evidence and falls through to rank 3. | | --scope | this.session | this.session | branch | ticket:<id> | session:<id> — where to recover from | | --conf-inconclusive | 0.60 | 0.0-1.0 — below this aggregate confidence => inconclusive.flag=true -> HITL | | --ladder | (full ladder) | comma-list to restrict sources (e.g. postflight_seed,ticket,ash_journal) | | --output | json | json (the envelope) | summary (human recap of the recovered intent) | | --dry-run | off | run RECON+SYNTHESIZE+HYPOTHESIZE, print the envelope, but do NOT hand it to any driver |

STOP-marker grammar (reused — emit exactly ONE as the last line when driven in a loop)

<!--ORCH-STATUS: STOP-DONE -->     envelope recovered + validated (confidence >= threshold)
<!--ORCH-STATUS: STOP-HITL -->     inconclusive.flag=true — ranked hypotheses attached for the operator
<!--ORCH-STATUS: STOP-ERROR -->    unreadable state / validator error
<!--ORCH-STATUS: CONTINUE -->      (when embedded in a loop) intent recovered; hand off to the next step

Protocol Rules (anti-fabrication invariants)

NO fabrication — every claim traces to a recovered_from source; ungrounded => inconclusive->HITL.
The envelope MUST pass validate_envelope.py before it is handed to any driver.
inconclusive.flag=true NEVER auto-drives a loop — it escalates with hypotheses.
Worktree discipline / never commit to main (this skill only READS + emits JSON; if it writes the envelope to disk, /tmp or the caller's scratch).
HUMAN_DOMAIN recovered as constraints -> guardrails; never actioned here.

DNA Geracional (inherited by every spawned sub-agent)

Dogfood: recover THIS session's own goal as the first validation fixture.
Persist-over-fail: the envelope IS the write-ahead checkpoint of the recovered intent.
DRY / KIS / YAGNI / SSOT — compose postflight+continuation-seed; never duplicate.
No self-destructive decisions; Boy-Scout — leave no fabricated intent behind.

Examples

goal-recovery                                   # recover this session's intent -> handoff-as-prompt
goal-recovery --scope=ticket:VKS-1234           # recover from a ticket anchor
goal-recovery --output=summary                  # human recap instead of the JSON envelope
goal-recovery --conf-inconclusive=0.75 --dry-run  # stricter HITL gate, print-only
goal-recovery --ladder=postflight_seed,ash_journal  # restrict the inference sources

Quality Tests (6/6 self-validity — dogfooded)

Self-Application — goal-recovery could recover its OWN creation goal ("realize the goal-loop contract's recovery step") from this session's state. PASS.
Non-Contradiction — composes (not duplicates) postflight P2/continuation-seed/pulse/preflight; the recovery + uncertainty-model + envelope are net-new. PASS.
Survival — applied to itself it advocates grounded, uncertainty-aware recovery; it is itself grounded (inference ladder + validator) and uncertainty-aware. PASS.
Bounded-Responsibility — single responsibility (recover -> typed envelope); --conf-inconclusive HITL gate; validator gate; no loop-driving here. PASS.
Explicit-Exception — When-not-to-use (S2 stated-goal) + inconclusive->HITL + §0 BEING>Rules + HUMAN_DOMAIN->guardrails. PASS.
Utility-Sunset — §DUED below. PASS.

scope-discipline 6Q: 6/6 (WHERE=multi-agent-os · DRY=gap-confirmed grep-absent, ~70% reused · WHY=recurring manual contract, this session = a Triple-touch instance · WHO=operator+amnesic agents · FITS=session-lifecycle sibling of preflight/postflight/pulse · MIN=1 skill + 1 schema). anti-theater 8Q REALITY: 8/8 (real gap · not theater · grounded ladder · not-invented · viable · applicable · implementable · useful).

§DUED Sunset (qualitative — not counter-based)

Deprecate when ANY: a host provides native session-intent recovery (E1) · agents gain reliable cross-session recall so an unstated goal never needs recovery (E3) · the family absorbs recovery into pulse/preflight (E6) · operator retraction (E4) · >=3 false recoveries where the ladder misfires (E5 -> refine, not auto-deprecate). Dormant-by-design otherwise.

templates/handoff-as-prompt.schema.json — the OUTPUT contract (extends continuation-seed)
templates/handoff-as-prompt.example.json — a valid fixture (this session's recovered intent)
bin/validate_envelope.py — the deterministic envelope validator + idempotency digest
skills/postflight/SKILL.md (+ templates/continuation-seed.template.json) — the forward-direction twin (reused)
skills/preflight/SKILL.md — R0 ticket-anchor (a ladder source)
skills/pulse/SKILL.md — state refresh (RECON) · skills/work-compass/SKILL.md — CPT N-Tree (a ladder source)
skills/agentic-session-harness/SKILL.md — journal (goal, spec_alignment) audit source
skills/ooda-loop/SKILL.md — the conductor that consumes this as its Observe step
skills/decompose-abstract-to-measurable/SKILL.md (Prisma) — consumes the recovered goal to derive the DoD
External grounding: GOOD (arXiv:2508.15119, uncertainty-aware goal inference) · RECAP (2509.04472, intent-rewrite) · Boyd OODA (Observe) · Raghavan & Schneier IEEE S&P 2025 (auditable Observe inputs)

Versioning

v0.2.0 (2026-08-14) — MINOR: braindump added to the inference ladder (rank 2) — a raw operator dump supplied AS the invocation subject, ranked above ticket because it is the operator's own words about this work, with an explicit drop-session-meta clause. Additive: no existing source changed and the envelope schema is unchanged, but recovery PRECEDENCE changes for any invocation that supplies a braindump: it now outranks ticket, objectives-N-Tree, ASH-journal, git and transcript. Ranks 2-7 shift to 3-8. handoff-as-prompt.schema.json gains braindump in the recovered_from[].source enum. Enables skills/refine-braindump-to-prompt PHASE 1 without forking a second recovery engine.
v0.1.0 (2026-07-12) — bootstrap. Recovery of an unstated session intent from an inference ladder; uncertainty-aware (ranked hypotheses + confidence + inconclusive->HITL); typed handoff-as-prompt envelope extending continuation-seed; direction-inverted twin of postflight; deterministic envelope over probabilistic content. Composes postflight/preflight/pulse/ASH; reimplements nothing.

License

MIT (matches multi-agent-os repo LICENSE).

Agent Skills: Goal-Recovery

Install this agent skill to your local

Skill Files