devrites 4.0.4 → 4.0.6

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 All notable changes to DevRites are documented here. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and DevRites adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). Releases are generated automatically by [semantic-release](https://semantic-release.gitbook.io/) from Conventional Commits on `main`.
 
+## [4.0.6](https://github.com/ViktorsBaikers/DevRites/compare/v4.0.5...v4.0.6) (2026-08-07)
+
+### Fixed
+
+* **rite:** gate one-shot evidence retention ([621a549](https://github.com/ViktorsBaikers/DevRites/commit/621a549acdf9ec2a0d427b829b0df8852502380d))
+
+## [4.0.5](https://github.com/ViktorsBaikers/DevRites/compare/v4.0.4...v4.0.5) (2026-08-07)
+
+### Fixed
+
+* **rite:** make recovery progress-aware ([87ecd07](https://github.com/ViktorsBaikers/DevRites/commit/87ecd07d406d8c9eed97f2d41ec3449b408e75bd))
+
 ## [4.0.4](https://github.com/ViktorsBaikers/DevRites/compare/v4.0.3...v4.0.4) (2026-08-07)
 
 ### Fixed

package/README.md

@@ -28,7 +28,7 @@ project-conventional push, tag, or PR action, and archive the workspace.
 Unattended runs may create local WIP checkpoint commits along the way, but they
 remain local unless Ship's disclosed plan includes an approved remote action.
 
-**Status:** [`v4.0.4`](https://github.com/ViktorsBaikers/DevRites/releases/tag/v4.0.4): see [`CHANGELOG.md`](CHANGELOG.md) for release notes.
+**Status:** [`v4.0.6`](https://github.com/ViktorsBaikers/DevRites/releases/tag/v4.0.6): see [`CHANGELOG.md`](CHANGELOG.md) for release notes.
 
 This is the latest published release; `main` may contain unreleased work.
 

package/pack/.claude/agents/devrites-plan-reviewer.md

@@ -55,7 +55,11 @@ line, and then assign the band. Do not choose a score and justify it afterward:
    model changes conservatively. Every destructive step needs a rollback.
 7. **Failure-mode coverage:** for each new codepath, find a realistic failure such
    as a timeout, nil value, race, or stale state. A silent failure with **no test AND
-   no error handling** is a critical gap.
+   no error handling** is a critical gap. For every consumptive action under
+   `one-shot-actions.md`, require durable bounded trust-safe evidence for every
+   terminal path plus unknown-well-formed, malformed/hostile, and cleanup-survival
+   fixtures. Missing evidence completeness is `broken`: the plan must not consume
+   the action to discover diagnostics that cleanup can erase.
 
 ## Confidence calibration + verification gate (mandatory)
 Give every finding a **confidence score from 1 to 10** and a quoted source:
@@ -73,7 +77,8 @@ Band each dimension `strong` / `adequate` / `thin` / `broken` (`broken` means
 Critical; `thin` means Important). For a borderline dimension, sample twice and
 take the **lower** band. The verdict uses the weakest dimension, not an average.
 Pass only when every dimension is at least `adequate`, no critical failure-mode gap
-remains, the shared-contract check passes, and the ID-and-meaning map contains no orphaned
+remains, every consumptive action passes one-shot evidence completeness, the
+shared-contract check passes, and the ID-and-meaning map contains no orphaned
 criterion, slice, or proof.
 
 ## Rules

package/pack/.claude/agents/devrites-slice-wright.md

@@ -97,10 +97,11 @@ an escalation and do not proceed.**
    not a quiet edit. The root inspects the returned diff and dedicated test
    analysis treats a weakened test as a Critical STOP.
    For non-trivial failure, invoke `devrites-debug-recovery` with the exact error,
-   hypotheses, and dead ends. Caller and recovery share three total failed attempts per
-   causal fingerprint. Count attempts supplied in the task/current context, include each
-   failure in `dead_ends`, and never make a fourth related attempt. At the limit return the
-   gate and repro. `Escalation` is only for product/irreversible choices
+   hypotheses, and dead ends. Caller and recovery share three no-progress attempts per
+   exact causal fingerprint. Count only corrections whose recheck preserves the same
+   decisive failure, using attempts supplied in the task/current context; include each in
+   `dead_ends`. Resolution is progress, while a different cause returns to the caller as a
+   new fingerprint. At the limit return the gate and repro. `Escalation` is only for product/irreversible choices
    or human-only access; technical failure is a blocker.
 5. **RETURN** the structured artifact (below) and stop. Do not start the next slice.
 

package/pack/.claude/skills/devrites-debug-recovery/SKILL.md

@@ -18,10 +18,14 @@ has no clear next move.
 1. **Build the feedback loop:** create a fast, deterministic, agent-runnable pass/fail
    signal. Spend most of the investigation here.
    See [build-the-loop.md](reference/build-the-loop.md).
-2. **Reproduce:** run the loop. Confirm the failure matches the user's report
-   (not a nearby failure); capture the **exact error text**; confirm
-   reproducibility (or a high enough repro rate for flaky bugs). Do not proceed
-   without reproduction.
+2. **Reproduce:** run the loop for a repeatable action. Confirm the failure matches
+   the user's report (not a nearby failure); capture the **exact error text**;
+   confirm reproducibility (or a high enough repro rate for flaky bugs). For a
+   consumptive action under
+   [`one-shot-actions.md`](../devrites-lib/reference/standards/one-shot-actions.md),
+   the retained bounded artifact
+   is the reproduction input and the action MUST NOT be rerun during diagnosis.
+   Do not proceed without one of those reproduction inputs.
 3. **Ranked hypotheses (3-5, falsifiable):** generate the list before testing
    any of them. Each must state a prediction.
    **Completion:** 3-5 distinct hypotheses each state an observable prediction.
@@ -51,21 +55,27 @@ has no clear next move.
 - **Do NOT loosen / delete a failing assertion** to get green: check whether
   it's drift first (route via `/rite-plan repair`).
 - **Do NOT hide flakiness** with sleeps / retries: characterize it.
-- **Re-run the original loop after the fix.** The minimized regression test is not
-  enough; prove the user-visible failure no longer reproduces.
+- **Re-run the original loop after the fix when it is repeatable.** For a
+  consumptive action, first re-vet evidence completeness and obtain any required
+  fresh authorization; offline fixtures remain mandatory but cannot authorize the
+  real attempt.
 - **Classify before routing** with
   [cleanup-and-classify.md](reference/cleanup-and-classify.md).
 - **Durably record class and rationale** in `decisions.md` and the applicable
   `evidence.md` or `## Dead ends` entry.
 - **One causal fingerprint, counted by the caller.** Normalize the root cause as
-  `<affected boundary>: <failure mechanism>` rather than hashing symptom text.
+  `<affected boundary>: <failed invariant/failure mechanism>` and bind its minimal
+  reproduction plus decisive signal rather than hashing symptom text.
   The caller and recovery attempts share one count: read the current context and
-  recorded `## Dead ends` / `evidence.md`, then include every failed attempt
+  recorded `## Dead ends` / `evidence.md`, then include every no-progress attempt
   with that fingerprint. Reclassify only on new causal evidence.
-- **A maximum of three total failed attempts per causal fingerprint stops the loop.** Record
-  attempt number, exact failure, hypothesis, probe, and failed idea after each
-  failure, and never make a fourth related attempt. There is no JSONL ledger,
+- **A maximum of three no-progress attempts per exact causal fingerprint stops the loop.**
+  Count an attempt only when its recheck preserves the same decisive failure.
+  Record attempt number, exact failure, hypothesis, probe, and failed idea after
+  each; closure is progress and a different Critical/Important invariant is a new
+  fingerprint. There is no JSONL ledger,
   counter command, or reset-on-green operation. Product/acceptance ambiguity, irreversible risk, or
   human-only access becomes a human gate; otherwise return reproducible `blocked` with
-  `Next: /rite-plan unblock`, never request attempt four. Coupled failure requiring behavior
-  change routes `/rite-plan repair`; behavior-neutral rerouting uses `unblock`.
+  `Next: none — technical recovery exhausted for <causal fingerprint>`, never request
+  attempt four. While budget remains, coupled failure requiring behavior change routes
+  `/rite-plan repair` inline; behavior-neutral rerouting uses `unblock` inline.

package/pack/.claude/skills/devrites-debug-recovery/reference/cleanup-and-classify.md

@@ -14,20 +14,25 @@ record the dead end and classify anew. Failure alone never resets the budget.
 - `preexisting`: the same failure exists outside the candidate delta. Record the baseline and fix it only when it blocks acceptance.
 - `not_a_defect`: the observation matches current accepted authority. Record that authority and continue.
 
-Only human credentials/quotas/actions or irreversible work pause; never ask to retry.
+Only human credentials/quotas/actions, irreversible work, or fresh authorization
+required by a consumptive action pause; never ask for a blind retry.
 
 Record class/routing in `decisions.md` and each failed attempt in `evidence.md`
 or `## Dead ends`. Use one stable causal fingerprint shaped as `<affected
-boundary>: <failure mechanism>`, not a hash of the symptom. Before retrying,
-count matching caller and recovery failures in the current context and those
-records. The maximum is three total failed attempts; never run a fourth related
-attempt. Also record baseline, exact failure, hypothesis/probe, attempt number,
-and any human predicate. Green evidence records the resolution but does not
-erase prior failures. There is no recovery counter file or command.
+boundary>: <failed invariant/failure mechanism>` and bind its minimal
+reproduction plus decisive signal, not a hash of the symptom. Before retrying,
+count matching caller and recovery no-progress attempts in the current context
+and those records. The maximum is three per exact fingerprint; count only a
+recheck that preserves the same decisive failure. Also record baseline, exact
+failure, hypothesis/probe, attempt number, and any human predicate. Green evidence
+records progress and resolution but does not erase prior failures. A different
+Critical/Important invariant is a new fingerprint. There is no recovery counter
+file or command.
 
 ## Cleanup checklist: required before declaring done
 
-- [ ] Original repro no longer reproduces (re-run the Phase 1 loop).
+- [ ] Original repeatable repro no longer reproduces, or the consumptive action's
+      offline regression and evidence-completeness fixtures pass without a rerun.
 - [ ] Regression test passes (or absence of seam is documented).
 - [ ] All `[DEBUG-...]` instrumentation removed (`grep` the prefix).
 - [ ] Throwaway harnesses deleted (or moved to a clearly marked debug location).

package/pack/.claude/skills/devrites-lib/reference/standards/README.md

@@ -36,6 +36,7 @@ topic's owner.
 | `context-hygiene.md` | Choosing `/clear`, `/compact`, or a handoff. |
 | `anti-patterns.md` | A pack-wide rationalization or red flag appears. |
 | `afk-hitl.md` | A pause, question, resume, or AFK decision is possible. |
+| `one-shot-actions.md` | A proof/action may be attempted once, needs fresh retry authorization, consumes external state/quota, or can delete its own failure evidence. |
 | `tooling.md` | Structural lookup, current external facts, or architecture memory is needed. |
 | `skill-authoring.md` | Creating, editing, routing, evaluating, or pruning a DevRites skill. |
 | `definition-of-done.md` | Prove, Seal, Ship, or Quick must decide whether work is finished. |

package/pack/.claude/skills/devrites-lib/reference/standards/afk-hitl.md

@@ -233,12 +233,22 @@ changes, not a request for permission to retry.
 
 ## Retry cap, no-progress loops, and self-resolve
 
-- **Cap retries:** three total attempts per root cause across wright and recovery is a hard
-  cap. The caller and recovery loop count every failed attempt with the same
-  causal fingerprint from the current context and recorded `## Dead ends` /
-  `evidence.md`. Persist the exact failure and attempted idea after each; never
-  make a fourth unchanged or related attempt for the same root cause. There is
-  no recovery counter file or command.
+- **Fingerprint the failed invariant, not the review round.** An exact causal
+  fingerprint is the owning gate/invariant plus its minimal reproduction and
+  decisive failure signal. A new DEC/DRIFT number, changed line number, reviewer
+  wording, or splitting one finding does not create a new fingerprint.
+- **Cap no-progress retries:** three no-progress attempts per exact causal fingerprint
+  across wright and recovery is a hard cap. Only a repair whose narrow recheck
+  leaves that fingerprint open or reproduces the same decisive failure consumes
+  an attempt. Closing a prior finding with discriminating evidence is progress:
+  mark that fingerprint resolved and do not charge it as no-progress. A genuinely
+  new Critical or Important finding with a different failed invariant and exact
+  evidence gets its own fingerprint and budget; Suggestion, Nit, FYI, or renamed
+  evidence cannot open, reset, or extend recovery.
+- **Persist accounting in existing records.** Record each fingerprint,
+  reproduction, attempted correction, `progress: resolved|no-progress`, and
+  decisive result in `drift.md` and `evidence.md`. Cold resume derives the count
+  from those records. There is no recovery counter file or command.
 - **Classify exhaustion:** human-owned contract/risk/access gaps open their gate. Otherwise
   preserve reproduction/dead ends, set `Status: blocked` and `Next step: none — technical recovery exhausted for <causal fingerprint>; requires new evidence or changed failure conditions`.
   Do not emit `/rite-plan unblock`, another phase command, a question, or

package/pack/.claude/skills/devrites-lib/reference/standards/one-shot-actions.md

@@ -0,0 +1,46 @@
+# One-shot evidence completeness
+
+An action is **consumptive** when a failed attempt is not safely equivalent to a
+normal rerun. This includes commands limited to one attempt, commands whose retry
+needs fresh human authorization, actions that spend external quota or mutate
+privileged/external state so a rerun is not equivalent, and actions whose cleanup
+can destroy the failure state needed for diagnosis. Successful cleanup does not
+make a consumptive action repeatable.
+
+## Pre-attempt gate
+
+Before Vet can emit READY, and again immediately before Prove executes the action,
+the approved `test-plan.md` must bind all of the following:
+
+1. **Durable retention:** an operator-controlled evidence artifact outside the
+   disposable runtime/cleanup tree, created before the first side effect, written
+   durably before cleanup, least-privilege, and bounded by schema, size, and
+   cardinality.
+2. **Trust-safe diagnostics:** known semantic values use the normal validator;
+   unknown but lexically well-formed non-secret values survive in bounded sanitized
+   fields; malformed, hostile, or secret-bearing values become fixed reason codes
+   rather than retained raw input.
+3. **Terminal completeness:** every success, nonzero exit, rejection, timeout,
+   signal, and cleanup failure either names the retained artifact or proves that
+   no diagnostic state exists. Failure retention preserves the original safe
+   failure family and cause through clean convergence.
+4. **Discriminating proof:** fixtures cover success, a known failure, an unknown
+   well-formed failure, malformed/hostile input, and cleanup after failure. They
+   prove cleanup cannot delete or overwrite the retained failure evidence.
+5. **Recovery sufficiency:** the retained bounded evidence is enough to choose an
+   offline correction or a truthful terminal classification without consuming
+   another attempt.
+
+Missing or stale evidence is an agent-owned technical plan gap: Vet returns
+`NEEDS REPLAN`, and Prove returns to Vet inline without executing the action. Never
+weaken the trust validator or spend the attempt merely to discover what the
+retention design should have preserved.
+
+## Failure handling
+
+After a consumptive action fails, its retained artifact is the reproduction input.
+Do not rerun the action during triage. A new attempt is admissible only after the
+evidence identifies a changed invariant or external condition, the affected plan
+and fixtures are re-vetted, and any required fresh authorization is obtained. If
+the attempt ran without complete retention, stop with the observed technical
+blocker; a blind retry cannot manufacture the destroyed evidence.

package/pack/.claude/skills/rite-autocomplete/reference/loop.md

@@ -59,11 +59,19 @@ Autocomplete root remains the caller:
 
 1. Save the originating phase/action in the native return cursor unless a valid
    one already exists.
-2. Invoke the required repair, Vet, remediation, and proof skills inline. Their
+2. Invoke the required repair, Vet, remediation, and proof skills inline. After
+   Plan repair, run Vet in its recovery mode as a narrow Vet recheck of the prior
+   findings, changed paths/criteria, and affected evidence; do not restart a Full
+   Vet over unchanged axes. Their
    `STOP` instructions end only those nested phases.
 3. Re-read `state.md` after each nested phase and follow its intermediate
    `next_action`; do not hand the intermediate command to the user.
-4. When the prerequisite chain is green, restore and consume the return cursor,
+4. Reconcile each exact causal fingerprint from `drift.md` and `evidence.md`.
+   Continue automatically when the recheck closes a prior fingerprint or admits
+   a genuinely new Critical/Important fingerprint. Charge only a no-progress
+   outcome against the same fingerprint; lower-severity novelty cannot prolong
+   the chain.
+5. When the prerequisite chain is green, restore and consume the return cursor,
    resume the originating phase, and continue the forward table.
 
 Count failed corrections by causal fingerprint under `afk-hitl.md`. Ask only

package/pack/.claude/skills/rite-autocomplete/reference/stop-conditions.md

@@ -5,6 +5,9 @@ the reason and one resume command when an actual human/safety/access action can
 change the state, notify if configured, and stop. Exhausted agent-owned
 technical recovery is terminal for its unchanged causal fingerprint: record
 `Next step: none` and no runnable recovery command. `--ship` cannot bypass them.
+Closure of a prior fingerprint is progress, not exhaustion. A separately evidenced
+Critical/Important failed invariant starts its own bounded fingerprint; it never
+resets or extends the budget of the one just closed.
 
 ## Always stop (irreversible-risk list: from `afk-hitl.md`)
 
@@ -24,8 +27,10 @@ Agent-owned backtracking is not a stop condition while its causal-fingerprint
 budget remains. The active caller invokes the earlier phase inline, follows Vet
 and any bounded remediation, then resumes the originating phase. Persist
 `Next step` for crash recovery, but do not surface it as a command the human must
-submit. Stop only when recovery is exhausted or the remaining choice is a real
-human/safety/access gate.
+submit. Use the repaired finding's narrow Vet recheck to distinguish resolution,
+the same decisive failure, and a genuinely new Critical/Important invariant.
+Stop only after three no-progress attempts on the exact same fingerprint or when
+the remaining choice is a real human/safety/access gate.
 
 On technical exhaustion, preserve the fingerprint, reproduction, attempts, and
 dead ends, then stop without `/rite-plan unblock` or another phase command.

package/pack/.claude/skills/rite-build/reference/afk-discipline.md

@@ -71,7 +71,9 @@ tests, types, or lint are red:
 The fail-on-red path:
 
 1. Continue the same wright under `devrites-debug-recovery`, carrying exact output and dead
-   ends; cap writer + recovery at three total attempts per root cause.
+   ends; cap writer + recovery at three no-progress attempts per exact causal fingerprint.
+   A correction that closes the reproduction is progress; a different evidenced
+   Critical/Important invariant starts a separate fingerprint.
 2. Green → record the slice. Product-contract/irreversible ambiguity → write the genuine
    human gate. Missing human-only credential/permission → write a human-intervention gate.
 3. Any other exhausted objective failure → set `Status: blocked`, preserve the reproduction,

package/pack/.claude/skills/rite-build/reference/phase-contract.md

@@ -33,8 +33,10 @@ See [`one-slice-cycle.md`](one-slice-cycle.md); candidate lifecycle is
    is not success.
 
    Technical failure uses bounded `devrites-debug-recovery`: reapply the host gate
-   before an accepted in-slice correction, keep one root cause/three total attempts,
-   and never rerun an unchanged check. Ask humans only for product decisions,
+   before an accepted in-slice correction, keep one exact fingerprint/three
+   no-progress attempts, and never rerun an unchanged check. A closed reproduction
+   is progress; a different Critical/Important invariant gets its own fingerprint.
+   Ask humans only for product decisions,
    irreversible risk, or genuinely human-only access.
 7. **Record.** After green proof, upsert `touched-files.md`'s authoritative candidate
    manifest from the actual scoped diff with explicit `present`/`deleted` rows;

package/pack/.claude/skills/rite-prove/SKILL.md

@@ -38,6 +38,8 @@ Pull these via `Read` when relevant:
   webhook / config / error messages / getting-started): **measure** the DX scorecard (run the flow,
   measure time-to-hello-world, and capture verbatim error text), rather than asserting it.
 - `definition-of-done.md`: acceptance, proof, gates, scope, rollback/docs.
+- `one-shot-actions.md`: pre-attempt evidence completeness and no-rerun handling for
+  consumptive proof actions.
 
 
 ## Operating rules
@@ -60,7 +62,10 @@ Pull these via `Read` when relevant:
 - **Prove remains the controlling caller during technical backtracking.** Save
   its return cursor, invoke Plan/Vet or bounded remediation inline, consume each
   nested phase boundary, then resume the failed Prove step. Never make the human
-  submit an agent-owned repair, re-vet, or proof-rerun command.
+  submit an agent-owned repair, re-vet, or proof-rerun command. Apply `afk-hitl.md`
+  progress accounting: only three no-progress attempts on the exact same fingerprint
+  exhaust recovery; a closed prior finding or a genuinely new
+  Critical/Important fingerprint continues inside this invocation.
 
 ## Released-workspace refresh entry
 
@@ -104,6 +109,13 @@ failure is a blocker; this Upgrade admission never authorizes source or test cha
    substituted commands, malformed manifests, zero-test/skipped/filtered results used as
    behavioral proof, success inferred only from exit status, and any source drift. Static
    gates prove only their named static criterion.
+   Immediately before any consumptive action, apply `one-shot-actions.md` to the
+   live candidate: require the vetted retained-artifact identity, bounds,
+   sanitization, terminal-path fixtures, and cleanup-survival proof. A missing,
+   stale, or disposable-only evidence surface returns to Vet inline and consumes
+   no attempt. Record the admitted artifact identity before execution. After a
+   failed consumptive action, triage from that artifact; never reproduce it by
+   rerunning the action.
 4. **UI feature?** The root applies the browser proof ladder with
    `design-brief.md`, `references.md`, the requested routes, browser harness, and
    allowed scratch path:

package/pack/.claude/skills/rite-prove/reference/failure-triage.md

@@ -4,7 +4,10 @@ When a test/build/runtime/browser check fails, triage before fixing. Delegates t
 loop to `devrites-debug-recovery`.
 
 ## Triage ladder
-1. **Reproduce:** run the failing command again; capture the exact error (quote it).
+1. **Reproduce:** for a repeatable command, run it again and capture the exact
+   error (quote it). For a consumptive action under
+   [`one-shot-actions.md`](../../devrites-lib/reference/standards/one-shot-actions.md),
+   never rerun it: use the retained bounded artifact as the reproduction input.
 2. **Classify** the failure:
    - test is right, code is wrong → fix the code (in scope).
    - test is wrong/outdated → that may be **spec drift** (acceptance criteria wrong).
@@ -14,13 +17,17 @@ loop to `devrites-debug-recovery`.
 3. **Isolate:** minimize to the smallest failing case; bisect the diff if needed.
 4. **Fix within scope:** the current slice/feature only. If the fix reaches outside
    scope, stop and record a blocker in `state.md` (then `/rite-plan unblock`).
-5. **Re-run** the same command; confirm green; record both attempts in `evidence.md`.
+5. **Re-run** the same command when it is repeatable; confirm green and record both
+   attempts in `evidence.md`. A consumptive action needs re-vetted evidence
+   completeness and any fresh authorization before a new attempt.
 
 ## Rules
 - Quote the real error text; don't paraphrase it away.
 - Don't loosen/delete a failing assertion to get green: investigate whether it's drift.
 - Don't add blanket retries/sleeps to hide flakiness.
-- Three failed fix attempts on the same root cause consume the shared recovery
-  cap: preserve the reproduction and dead ends, then stop once as a technical
-  blocker. Ask only when the remaining decision is human-owned.
+- Three no-progress attempts on the exact same causal fingerprint consume the
+  shared recovery cap. A recheck that closes the reproduction is progress; a
+  different evidenced Critical/Important invariant gets its own fingerprint.
+  Preserve the reproduction and dead ends, then stop once as a technical blocker.
+  Ask only when the remaining decision is human-owned.
 - A failure you can't fix in scope is a recorded **blocker**, not a silent skip.

package/pack/.claude/skills/rite-vet/SKILL.md

@@ -20,7 +20,8 @@ but every profile keeps the exact plan-reviewer gate.
 Pull the standard named by the active axis: `principles.md`, `patterns.md`,
 `coding-style.md`, `testing.md`, `spec-grammar.md`, `performance.md`,
 `error-handling.md`, `development-workflow.md`, `afk-hitl.md`,
-`developer-experience.md`, `elicitation.md`, and `definition-of-done.md`.
+`one-shot-actions.md`, `developer-experience.md`, `elicitation.md`, and
+`definition-of-done.md`.
 
 
 ## Operating rules
@@ -43,6 +44,9 @@ Pull the standard named by the active axis: `principles.md`, `patterns.md`,
 - **Honor a recovery origin.** Preserve a valid technical-backtracking return
   cursor throughout review. Agent-owned `NEEDS REPLAN` returns internally to the
   controlling caller; it never becomes a request for the human to invoke Plan.
+- **Recovery recheck is bounded.** A valid technical-backtracking cursor plus a
+  recorded open fingerprint enters the Recovery recheck below. Recovery recheck
+  does not start another Full Vet or repeat unaffected axes/reviewers.
 
 ## Workflow
 0. **Read `.claude/skills/devrites-lib/reference/standards/core.md`** first.
@@ -57,12 +61,26 @@ Pull the standard named by the active axis: `principles.md`, `patterns.md`,
    `.claude/skills/devrites-lib/reference/standards/tooling.md`) for placement / blast-radius / reuse checks.
 1. **Set review depth. Never skip this step.** Apply
    [`reference/depth.md`](reference/depth.md) exactly. Every plan leaves a recorded
-   engineering verdict and `test-plan.md` coverage map.
-1a. **Independent pass at every depth.** Freeze the candidate; dispatch exact
+   engineering verdict and `test-plan.md` coverage map. For an admitted Recovery
+   recheck, retain the prior depth and proceed to step 1b instead of starting a new
+   depth/pass.
+1a. **Independent pass at every initial depth.** Freeze the candidate; dispatch exact
    `devrites-plan-reviewer` fresh/read-only and validate its report. Add exact
    `devrites-devex-reviewer` for developer surfaces and require the current exact
    `devrites-strategy-reviewer` verdict after significant Temper. Missing accounts
    block; never substitute inline work.
+1b. **Recovery recheck.** Require the valid return cursor, prior accepted findings,
+   exact open fingerprint/reproduction, repaired candidate identity, changed
+   paths/criteria, and affected evidence from `drift.md` / `evidence.md`. Freeze
+   that packet and dispatch each exact owning reviewer once, fresh/read-only and
+   limited to those inputs. Do not rerun unaffected reviewers or the broad steps
+   2-4 inventory. Mark the prior fingerprint resolved only when discriminating
+   evidence closes its reproduction; otherwise record one no-progress outcome.
+   A newly caused or exposed Critical or Important finding may block only with a
+   different failed invariant, exact evidence in the changed/affected boundary,
+   and a new fingerprint. A Suggestion, Nit, or FYI cannot keep recovery open.
+   Reconcile the common artifact/readiness gates below, then return to the caller
+   or its next repair; never convert the nested result into a human command.
 2. **Scope challenge (blocking gate):** apply §0 of
    [`reference/review-axes.md`](reference/review-axes.md). Search accepted ADRs and
    relevant workspace `decisions.md` files directly. Harden to the smallest
@@ -80,6 +98,11 @@ Pull the standard named by the active axis: `principles.md`, `patterns.md`,
    unmeasurable conflict = gap. Record complete SHA-256 provenance inputs. Require
    every behavioral mapping to name a positive,
    discriminating assertion and decisive signal, not merely a command or expected exit zero.
+   Identify every consumptive action under `one-shot-actions.md`. Before admitting
+   it, require the exact durable retention surface, trust-safe diagnostic schema,
+   cleanup ordering, terminal-path coverage, and discriminating fixtures in
+   `test-plan.md`. Missing or stale one-shot evidence completeness is a technical
+   preflight gap; do not spend the action to learn what cleanup would erase.
    Preflight observes; it need not make future behavior pass.
 2c. **Implementation-readiness audit.** Goal-backward map every REQ/AC/NFR, interaction,
    edge/prohibition, and decision-coverage row to a slice and executable proof. Verify

package/pack/.claude/skills/rite-vet/reference/artifacts.md

@@ -35,6 +35,7 @@ Inventory/currentness: <pass/gaps> · slice order/independence: <pass/gaps> ·
 UX/spec/architecture: <pass/n-a/gaps> · operations/rollout/rollback: <pass/n-a/gaps>
 
 Shared contract proof: <pass | gap: missing/one-sided/duplicated-contract/vague/non-consuming>
+One-shot evidence completeness: <n/a | pass: action + retained artifact + bounds/sanitization + cleanup-survival fixtures | gap: exact missing proof>
 
 ## 3. Axis findings (floor-gated)
 | Axis | Floor band | Findings (sev · confidence) |
@@ -94,6 +95,11 @@ Commands in this durable artifact are portable repository commands: no RTK or lo
 aliases, user-specific absolute paths, or temporary proof trees. Evidence records
 the command actually executed.
 
+## Consumptive action gates
+| Action | Why one-shot/consumptive | Retained artifact | Bounds + sanitization | Terminal-path fixtures | Cleanup survival | Retry authority |
+|---|---|---|---|---|---|---|
+| <exact approved action or n/a> | <one attempt/quota/state/evidence deletion> | <durable operator-controlled path/schema> | <size/cardinality + known/unknown/malformed handling> | <success/known/unknown/hostile> | <assertion> | <none/fresh human authorization> |
+
 Every behavioral row names a positive, discriminating assertion and the decisive output it
 produces. A command or expected exit zero alone is not a behavioral assertion; static gates
 prove only their named static criterion.

package/pack/generated/claude/agents/devrites-plan-reviewer.md

@@ -55,7 +55,11 @@ line, and then assign the band. Do not choose a score and justify it afterward:
    model changes conservatively. Every destructive step needs a rollback.
 7. **Failure-mode coverage:** for each new codepath, find a realistic failure such
    as a timeout, nil value, race, or stale state. A silent failure with **no test AND
-   no error handling** is a critical gap.
+   no error handling** is a critical gap. For every consumptive action under
+   `one-shot-actions.md`, require durable bounded trust-safe evidence for every
+   terminal path plus unknown-well-formed, malformed/hostile, and cleanup-survival
+   fixtures. Missing evidence completeness is `broken`: the plan must not consume
+   the action to discover diagnostics that cleanup can erase.
 
 ## Confidence calibration + verification gate (mandatory)
 Give every finding a **confidence score from 1 to 10** and a quoted source:
@@ -73,7 +77,8 @@ Band each dimension `strong` / `adequate` / `thin` / `broken` (`broken` means
 Critical; `thin` means Important). For a borderline dimension, sample twice and
 take the **lower** band. The verdict uses the weakest dimension, not an average.
 Pass only when every dimension is at least `adequate`, no critical failure-mode gap
-remains, the shared-contract check passes, and the ID-and-meaning map contains no orphaned
+remains, every consumptive action passes one-shot evidence completeness, the
+shared-contract check passes, and the ID-and-meaning map contains no orphaned
 criterion, slice, or proof.
 
 ## Rules

package/pack/generated/claude/agents/devrites-slice-wright.md

@@ -97,10 +97,11 @@ an escalation and do not proceed.**
    not a quiet edit. The root inspects the returned diff and dedicated test
    analysis treats a weakened test as a Critical STOP.
    For non-trivial failure, invoke `devrites-debug-recovery` with the exact error,
-   hypotheses, and dead ends. Caller and recovery share three total failed attempts per
-   causal fingerprint. Count attempts supplied in the task/current context, include each
-   failure in `dead_ends`, and never make a fourth related attempt. At the limit return the
-   gate and repro. `Escalation` is only for product/irreversible choices
+   hypotheses, and dead ends. Caller and recovery share three no-progress attempts per
+   exact causal fingerprint. Count only corrections whose recheck preserves the same
+   decisive failure, using attempts supplied in the task/current context; include each in
+   `dead_ends`. Resolution is progress, while a different cause returns to the caller as a
+   new fingerprint. At the limit return the gate and repro. `Escalation` is only for product/irreversible choices
    or human-only access; technical failure is a blocker.
 5. **RETURN** the structured artifact (below) and stop. Do not start the next slice.
 

package/pack/generated/claude/skills/devrites-debug-recovery/SKILL.md

@@ -18,10 +18,14 @@ has no clear next move.
 1. **Build the feedback loop:** create a fast, deterministic, agent-runnable pass/fail
    signal. Spend most of the investigation here.
    See [build-the-loop.md](reference/build-the-loop.md).
-2. **Reproduce:** run the loop. Confirm the failure matches the user's report
-   (not a nearby failure); capture the **exact error text**; confirm
-   reproducibility (or a high enough repro rate for flaky bugs). Do not proceed
-   without reproduction.
+2. **Reproduce:** run the loop for a repeatable action. Confirm the failure matches
+   the user's report (not a nearby failure); capture the **exact error text**;
+   confirm reproducibility (or a high enough repro rate for flaky bugs). For a
+   consumptive action under
+   [`one-shot-actions.md`](../devrites-lib/reference/standards/one-shot-actions.md),
+   the retained bounded artifact
+   is the reproduction input and the action MUST NOT be rerun during diagnosis.
+   Do not proceed without one of those reproduction inputs.
 3. **Ranked hypotheses (3-5, falsifiable):** generate the list before testing
    any of them. Each must state a prediction.
    **Completion:** 3-5 distinct hypotheses each state an observable prediction.
@@ -51,21 +55,27 @@ has no clear next move.
 - **Do NOT loosen / delete a failing assertion** to get green: check whether
   it's drift first (route via `/rite-plan repair`).
 - **Do NOT hide flakiness** with sleeps / retries: characterize it.
-- **Re-run the original loop after the fix.** The minimized regression test is not
-  enough; prove the user-visible failure no longer reproduces.
+- **Re-run the original loop after the fix when it is repeatable.** For a
+  consumptive action, first re-vet evidence completeness and obtain any required
+  fresh authorization; offline fixtures remain mandatory but cannot authorize the
+  real attempt.
 - **Classify before routing** with
   [cleanup-and-classify.md](reference/cleanup-and-classify.md).
 - **Durably record class and rationale** in `decisions.md` and the applicable
   `evidence.md` or `## Dead ends` entry.
 - **One causal fingerprint, counted by the caller.** Normalize the root cause as
-  `<affected boundary>: <failure mechanism>` rather than hashing symptom text.
+  `<affected boundary>: <failed invariant/failure mechanism>` and bind its minimal
+  reproduction plus decisive signal rather than hashing symptom text.
   The caller and recovery attempts share one count: read the current context and
-  recorded `## Dead ends` / `evidence.md`, then include every failed attempt
+  recorded `## Dead ends` / `evidence.md`, then include every no-progress attempt
   with that fingerprint. Reclassify only on new causal evidence.
-- **A maximum of three total failed attempts per causal fingerprint stops the loop.** Record
-  attempt number, exact failure, hypothesis, probe, and failed idea after each
-  failure, and never make a fourth related attempt. There is no JSONL ledger,
+- **A maximum of three no-progress attempts per exact causal fingerprint stops the loop.**
+  Count an attempt only when its recheck preserves the same decisive failure.
+  Record attempt number, exact failure, hypothesis, probe, and failed idea after
+  each; closure is progress and a different Critical/Important invariant is a new
+  fingerprint. There is no JSONL ledger,
   counter command, or reset-on-green operation. Product/acceptance ambiguity, irreversible risk, or
   human-only access becomes a human gate; otherwise return reproducible `blocked` with
-  `Next: /rite-plan unblock`, never request attempt four. Coupled failure requiring behavior
-  change routes `/rite-plan repair`; behavior-neutral rerouting uses `unblock`.
+  `Next: none — technical recovery exhausted for <causal fingerprint>`, never request
+  attempt four. While budget remains, coupled failure requiring behavior change routes
+  `/rite-plan repair` inline; behavior-neutral rerouting uses `unblock` inline.