Downstream trial: Hermes Maintenance package (2026-08-10)

Operator intent

Use the live Process Engine to create a Turnstone package from accumulated evidence of operating/integrating/maintaining/troubleshooting Hermes through its sanctioned Hermes MCP path: project Hermes Maintenance, persona Hermes Manager, skillset Hermes Management. Reuse existing evidence rather than making the operator repeat material; extract only what materially belongs; preserve provenance; distinguish current practice from superseded history.

Trial classification

Downstream product trial - targeted post-R1 live trial (first full resource/performance baseline). Hermes/agent management remains close to the infrastructure/admin evidence base; this trial does not claim broad cross-domain generalization.

Canonical and live state

Canonical PE SHA (start and end): 076dc4d323dba449cf416c1f1db948e997b0b334
Tag v1.9.6 → 65eaba888322609afa9986ff1d19e959317f5e28 (unchanged)
Live PE at trial start (05:49Z): ALIGNED - all six pipeline skills matched canonical
Live PE at trial end (06:28Z): UNEXPECTED DRIFT - all six reverted to v1.8-era content by the Shared Skills Repo Sync & Drift 1300-BKK schedule (13:02:39 BKK; mutation-ledger seq 363-368 db-skill-content-sync; updated timestamps 06:02:38/39Z match exactly). Recurrence of the historical same-ID content reversion; writer mechanism now identified (the sync writes stale shared-workspace/skills/ content into the DB when it deems DB parity content-real). Not repaired in this gate.

Package produced (draft - NOT shipped)

operations/hermes-maintenance-package/ - 17 files: manifest, persona Hermes Manager, prompt-policy, judge-rules, evidence-library, interface map, creation-checklist, trial-plan, PROCESS-REPORT, trial-evidence + 7 skills (parent hermes-management + 6 children: access-connectivity, health-observability, delegation, troubleshooting, update-maintenance, verification).

Context reuse (relevant-context test)

Sources used without re-asking: deployment facts YAML (Hermes/MCP/delegation), mutation ledger (Hermes entries 08-03→08-08), hermes_gateway_mcp.py adapter, Schedule 3 definition, stewardship playbook, chat-fix log, session memories, OpenClaw package as shape reference.
Superseded history correctly separated: 2026-07-02 stewardship playbook and deprecated cross-API runbook labeled as superseded, retained for provenance.
Operator questions asked: 2 (Summary Gate GO; manifest-retry choice). Both closed real gates; no re-asking of retrievable facts.

Review

PASS after 1 revision loop (unquoted colon in hermes-troubleshooting description → YAML parse failure → quoted → re-verified all 7).
Frontmatter parse, anatomy 8/8, sanitization, governance posture, allowed-fields all checked.

Trial

PASS (read-only + gated near-misses): T1 access/health PROVEN, T2 capabilities PROVEN, T3 detailed health PROVEN, T4 models PROVEN (sanitized), T5/T6/T7/T9/T10 gated near-miss (no side effect), T8 collision PROVEN non-destructive.
No mutation; manifest trial_evidence filled only from actual results.

R1 destination-skill behavior observed

Existing Hermes Management v1.0.0 skill discovered naturally via normal Collect (no hint).
Identity-binding observation: the generator queried get("Hermes Management") (existing skill's display name) and get("openclaw-management") (convention check) but did not run an exact-name get on the generated canonical name hermes-management - so no explicit PROVEN ABSENT receipt was produced for the new canonical identity. Collision handled conservatively (non-destructive, operator-gated). Candidate lesson: R1 evidence should be recorded for the generated canonical name even when a display-title collision exists.
Summary-gate fidelity observation: the generator presented the Summary Gate but began authoring without awaiting confirmation; it stopped only when the write approvals resolved as denied, then re-presented the gate. Candidate lesson: Summary Gate should be enforced as a real stop before authoring.

Telemetry (first full baseline)

Wall-clock trial duration: ~38 min (05:49:50Z → ~06:28Z / 12:49:50 → 13:28 BKK).
Model/provider: deepseek-v4-flash (deepseek-deepseek-v4-flash) - bundled deployment.
Context-window high-water: ~3.6M tokens (context_ratio 0.154) - occupancy, not consumed tokens.
Authoritative per-call token usage: UNAVAILABLE (provider API not exposed).
Model calls: not separately countable from session record (multi-turn with tool batches).
Cost: not computed - bundled/subscription runtime, no per-call pricing basis exposed.

Findings

PE layer: Summary-Gate enforcement; R1 canonical-name evidence completeness.
Generated package: high quality, MCP-only, evidence-state labeled, operator-gated; no creds/private hosts leaked.
Runtime: approval-path inconsistency (write approvals sometimes resolved as denied; always:true + explicit cycle resolve worked).
Runtime/MCP: hermes-gateway path healthy (v0.20.0); model-options default observed gpt-5.6-luna on openai-codex.
Drift recurrence: shared-skills sync reverted live PE during the trial; mechanism identified; containment decision required before live integration/ship.

Stop reason

Ship held per trial gate. Package is eligible for operator Ship gate - NOT deployed, NOT published, collision disposition + publish scope remain operator decisions.

Candidate engine lessons (not implemented)

Summary Gate must be a hard stop before authoring.
R1 destination evidence should be explicitly recorded for the generated canonical identity (exact-name get) in addition to any display-title collision inspection.
Live PE integration remains HELD until the shared-skills sync drift recurrence is contained.

Containment (operator-authorized)

Following the recurrence, the operator authorized updating/aligning the Shared Skills Repo Sync & Drift 1300-BKK scheduled task so it stops reverting live PE content from the stale shared source. This is a scheduled-task correction only; no broader PE reconciliation was performed in the trial gate.

Evidence-integrity audit - Hermes Maintenance trial (2026-08-10)

Read-only evidence reconstruction. No deployment, no PE prompt changes, no branch, no repair in this gate.

A. Trial vs post-trial separation

Trial window: 2026-08-10 05:49:59Z (WS 214aea26… created) → 06:34:02Z (WS closed). Product trial: Pattern → Review → Trial, ship held.
Post-trial (separately operator-authorized): sync-schedule disable, DB PE repair, shared-copy update, Operational-Commons v1.9.6 alignment corrections (advisory governance, spec-vs-guidance, triage traceability, validator fidelity, OWASP proportional) #6, llama-turbo-rebuild commit. All occurred after 06:34:02Z and are not part of the trial verdict. No post-trial action lacked an explicit operator authorization (each was directed in a subsequent operator message).

B. Mid-trial version timeline (stage-by-stage)

Governing content = core skill v27 (canonical 076dc4d) - pinned into the trial WS at creation by Turnstone snapshot isolation (applied_skill_content). The 06:02:38-39Z DB overwrite created v28 (reverted v1.8-era) in the DB, but the running session retained v27. No pipeline stage was governed by reverted content.

Stage	Executed	Governing skill	Version	Canonical?
Intent/Orient/Collect	05:49:59-05:56Z	core (applied)	v27	✅ canonical
Clarify/Objective	~05:56-06:00Z	core	v27	✅ canonical
Summary Gate + pre-GO authoring attempt	~06:00-06:02Z	core	v27	✅ canonical
DB overwrite (v28 reverted)	06:02:38Z	- (session unaffected)	-	-
Summary Gate re-presented + GO	~06:05-06:09Z	core	v27	✅ canonical
Pattern/Author	06:09-06:20Z	core (pattern-author skill NOT loaded)	v27	✅ canonical
Review	06:20-06:22Z	core (review skill NOT loaded)	v27	✅ canonical
Trial	06:21-06:23Z	core (trial skill NOT loaded)	v27	✅ canonical

Version binding: PROVEN - canonical 076dc4d core governed every stage. Pattern/review/trial skills were never skills load-ed (verified: 83 tool calls, 3 skills find/get, 0 skills load); the model emulated those stages from core routing + reference files + its own knowledge. Stage skill versions therefore: not applicable (not loaded); core version binding: PROVEN.

C. Summary Gate classification

msg[46] (~06:00Z): Summary Gate presented and 4 write_file calls issued in the same turn (manifest, persona, prompt-policy, judge-rules).
msg[47-50]: all 4 returned "Denied by user" (approval wall).
msg[54]: model self-corrected - "I jumped ahead: I presented the summary gate and then started authoring in the same turn…" and saved a feedback memory.
msg[58] re-presented gate; msg[59] operator GO; msg[60] authoring resumed.
No durable artifact was created before GO (all pre-GO writes denied).

Classification: CANONICAL CORE VIOLATION - behavioral adherence failure. Canonical v27 explicitly states "proceed only on confirmation"; the model presented the gate but began authoring in the same turn. The product spec is adequate; this is an execution/model-adherence gap, not a missing requirement.

D. R1 destination evidence - original receipts (per generated skill)

Original session catalog calls (msg indices): find("hermes") [8], get("Hermes Management") [15], get("openclaw-management") [60]; plus Hermes skills_list [121] (Hermes store, not Turnstone catalog).

Generated canonical name	Original exact-name query	Original state	Receipt bound to canonical name?
`hermes-management`	none (only `get("Hermes Management")` display title)	UNPROVEN (display-title overlap recorded)	❌ no
`hermes-access-connectivity`	none	UNPROVEN	❌ no
`hermes-health-observability`	none	UNPROVEN	❌ no
`hermes-delegation`	none	UNPROVEN	❌ no
`hermes-troubleshooting`	none	UNPROVEN	❌ no
`hermes-update-maintenance`	none	UNPROVEN	❌ no
`hermes-verification`	none	UNPROVEN	❌ no

Authoritative identity resolution: existing object = name='Hermes Management', template_id=4c72a8fd… (v1.0.0). Generated canonical = hermes-management. These are distinct identities under exact-name semantics; a query for Hermes Management is NOT evidence for hermes-management.

POST-TRIAL EVIDENCE (current read-only lookups, 2026-08-10): all 7 generated canonical names ABSENT from the catalog (exact-name). No actual identity collision exists; the overlap is display-title only.

E. Historical Review verdict reassessment

Native parse receipts: NOT produced (Review used self-authored bash python3 yaml checks, not the native Turnstone parse operation).
Allowed-field comparison: informal only (no separate native contract comparison).
Destination-state evidence per generated skill (canonical R1 requirement): absent for all 7 (no exact-name query at Review time).
Capability-path: reviewed against Hermes MCP path (adequate but emulated).
Parent collision disposition: explicit, reversible, operator-gated (coexist proposed) - but based on a different identity (display title).
Material identity evidence: UNPROVEN for all 7 canonical names at Review time.

Original Review PASS NOT SUPPORTED under the canonical R1/Review contract. The draft could receive a fresh canonical re-review later (read-only) without regeneration; that re-review is not performed in this gate.

F. Trial completeness

Package is activation-dependent (parent + 6 children with activation fields).
Trial plan T1-T10: operational cases (health, capabilities, models, gated near-misses). These are operational behavior cases, NOT activation/routing tests.
No should-trigger, shouldn't-trigger, parent activation, child-selection, near-match, or wrong-child cases were run.
Acceptance/happy path: covered (T1-T4 live read-only). Boundary/adversarial: partial (gated near-misses). Identity: partial (collision observation, but not per-canonical-name). Capability preflight: covered (T1-T4 live MCP). Safeguards: not risk-relevant, correctly minimal. Activation/routing: applicable but UNRUN.

Corrected overall Trial verdict: INCOMPLETE (activation/routing category applicable and unrun; INCOMPLETE does not advance to Ship).

G. Ship eligibility

Generated package quality: high (MCP-only, evidence-state labeled, operator-gated, no secrets leaked; filesystem-only draft).
Historical Review verdict: NOT SUPPORTED.
Historical Trial verdict: INCOMPLETE (activation unrun).
Current destination-state completeness: all 7 canonical names PROVEN ABSENT (post-trial, fresh evidence - cannot retroactively validate the original Review).
Current Ship eligibility: NOT SHIP-ELIGIBLE - RE-REVIEW + TRIAL COMPLETION REQUIRED.

H. Two PE lessons - classification

Summary Gate: ALREADY SPECIFIED / EXECUTION FAILURE. Canonical core v27 states "proceed only on confirmation"; the model attempted authoring pre-confirmation. Not a prompt gap.
R1 canonical-name evidence: ALREADY SPECIFIED / EXECUTION FAILURE (prompt-adherence). Canonical Pattern v27 requires destination-skill evidence bound to the generated canonical name; the model queried a display title instead. The spec text is unambiguous; the model failed to follow it. (Note: pattern-author was never loaded as a skill - the model emulated it from core + references; the failure is in emulation adherence, and stage-skill version binding is UNPROVEN for that emulation.)

Neither lesson is READY TO DEFINE as a product-prompt change from this trial alone.

I. Manifest classification + file count

Hermes manifest.json = local trial/deployment evidence artifact (mirrors OpenClaw package convention). Canonical PE explicitly says "no new manifest or persisted schema" as a product authority; this package-local record is consistent with that (local evidence only, not lifecycle authority). No architectural drift.
17-vs-18 discrepancy: genuine package = 17 files. The 18th entry is trial-ws-final.json - a stray monitoring artifact (Turnstone session-detail capture) that landed in the package directory during the trial session; it is not part of the package. Issue Downstream trial: Hermes Maintenance package (2026-08-10) #13's "17 files" was correct.

J. Corrected telemetry baseline

Wall-clock: 05:49:59Z → 06:34:02Z (≈44 min incl. operator waits; product pipeline ≈ 33 min to trial completion).
Context high-water ~3.6M tokens = context-window occupancy, not consumed tokens.
Authoritative consumed tokens: UNAVAILABLE (no provider API surface exposed).
Model calls: UNAVAILABLE (assistant turns ≈ 30 is a message-count proxy, not an authoritative model-call count).
Cost: UNAVAILABLE / bundled runtime, no attributable pricing basis.
Questions: 2 (Summary Gate GO; manifest-retry choice). Revision loops: 1 (unquoted-colon frontmatter fix).

K. Operational-Commons boundary

Operational-Commons #6 is the correct home for the shared-skills sync defect. No commits, no schedule re-enable, no H/O resync, no drift-checker/source-hierarchy changes performed in this gate. GitHub RedEyeNinja-BKK/Process-Engine remains PE product authority.

M. Current PE canonical/live

GitHub canonical: 076dc4d323dba449cf416c1f1db948e997b0b334 (verified).
Live PE: all six pipeline skills MATCH canonical (read-back verified). ALIGNED.
No new drift observed; sync schedule remains disabled (containment).

Conclusion

The Hermes Maintenance trial proves behavior of canonical 076dc4d core (version binding PROVEN via snapshot isolation). Review PASS and Trial PASS as originally recorded are not supported under the canonical contracts that governed the session: Review lacked required R1 destination-state evidence per generated skill and native parse receipts; Trial omitted the applicable activation/routing category. The two apparent PE findings are execution/adherence failures against already-adequate canonical text, not missing product requirements.

Canonical stage-bound re-review + Trial completion (2026-08-10, follow-up gate)

Historical findings preserved: canonical Core v27 snapshot-bound; Summary Gate adherence failure; Pattern/Review/Trial not loaded in the original run; original Review PASS NOT SUPPORTED; original Trial INCOMPLETE; package NOT SHIP-ELIGIBLE. This follow-up produces new canonical Review/Trial evidence, separately labeled.

Routing/load semantics (native)

Core route → process-engine-review is conceptual routing only - it names the stage but does NOT load the skill.
Native activation requires explicit skills(action='load', name=…) (or /skill, or spawn_workstream(skill=…)), which calls set_skill() → _load_skills() → replaces _skill_content.
Evidence a skill is active: persisted applied_skill_id / applied_skill_version / applied_skill_content + _skill_name / _skill_content.
EXPECTED: routed sub-skill must be natively loaded to govern. OBSERVED (original trial): Core loaded; pattern/review/trial NOT loaded - stages emulated from Core routing.

Fresh Review (loaded `process-engine-review`, WS `f34fda6b…` / `7b22dc71…` / `12a26530…`, skill_id `8f8411e8…`)

Review 1 verdict: REVISE (F1-F4). The loaded Review skill caught: F1 non-contract frontmatter keys (7/7), F2 no native parse receipts, F3 destination evidence misbound (parent receipt for Hermes Management ≠ hermes-management; 6 children UNPROVEN), F4 local-CLI ops with no declared executable path.
Pattern revision (loaded process-engine-pattern-author, WS 8101ffc9…): F1 fixed (version/kind/activation → metadata; allowed_tools → allowed-tools); F2 parse receipts 7/7 recorded; F3 fresh per-canonical-name preflight - hermes-management PROVEN PRESENT (display-identity collision, disposition = intentional coexistence + operator-gated reversible disable, canonical name unchanged), 6 children PROVEN ABSENT; F4 local-CLI path declared with guardrails (allowed-tools: local-cli + metadata.execution-paths).
Review 2 verdict: REVISE (F5 narrow). F1-F4 verified fixed; F5 = trial plan not reconciled with F4 path (stale harness rule, no local-CLI cases).
Pattern F5 fix (WS 005ff738…): T11-T13 scripted (read-only auto / mutation near-miss / read-back verify), harness two-path, coverage boundary marked (MCP PROVEN / local-cli PENDING), manifest status refreshed.
Review 3 verdict: PASS. F5 addressed, F1-F4 no regression, independently verified. Ship NOT approved - operator gate + collision disposition remain.

Fresh Trial (loaded `process-engine-trial`, WS `57419ea9…`, skill_id `ecf4387b…`)

Applicable categories determined before execution (acceptance/happy PROVEN via historical T1-T4; boundary/gray-zone PROVEN via T5-T10; identity PROVEN + new C-set; trigger/activation PROVEN this round; safeguards PARTIAL - local-cli T11-T13 PENDING).
Activation/routing suite: 37/37 PASS - should-trigger recall 12/12; shouldn't-trigger precision 11/11; identity/near-match/wrong-child 10/10 (no entity collapse, Hermes Management skill-update correctly NONE+CLARIFY); ambiguous clarification 4/4; blind second evaluator 37/37 identical.
Historical T1-T10 receipts preserved as prior direct evidence (not re-run for ceremony).
Overall Trial verdict: INCOMPLETE - every executed category PASSES, but local-cli surface T11-T13 remains PENDING (blocked by missing deploy-time local-cli mapping, an operator ship-time decision). INCOMPLETE does not advance to Ship.

Stage-load diagnostics

Review: STAGE-LOAD ROOT CAUSE SUPPORTED - the loaded Review skill immediately required all-seven R1 evidence, native parse receipts, allowed-field separation, wrong-entity analysis, and capability-path evidence; the emulated Review missed all of these.
Trial: STAGE-LOAD ROOT CAUSE SUPPORTED - the loaded Trial skill identified activation/routing as applicable before execution and prevented an overall PASS while the category was missing (issued INCOMPLETE); the emulated Trial gave PASS with no trigger suite.

Routing/load product classification

ROUTING CONTRACT AMBIGUITY - READY TO DEFINE (candidate only). Core names/routes to stage skills but does not make the required Turnstone-native activation/load step explicit; explicit loading demonstrably restored the missing stage contracts. NOT implemented in this gate; no PE patch made.

Summary Gate classification (unchanged)

ALREADY SPECIFIED / BEHAVIORAL ADHERENCE FAILURE - canonical Core active, said "proceed only on confirmation"; model attempted writes before GO; approval wall prevented durable mutation. No prompt strengthening from this occurrence alone.

Follow-up resource telemetry

Start 07:42:04Z → end ~09:05Z (≈83 min; operator approval/waiting cycles interspersed; model deepseek-v4-flash bundled).
Authoritative consumed tokens: UNAVAILABLE; context high-water not separately captured per WS; cost: UNAVAILABLE / bundled runtime.
Revision loops in this gate: 2 (F1-F4 → F5). Operator decisions required: none new (collision disposition + local-cli mapping remain open ship-time decisions).

Current alignment

GitHub 076dc4d… = live six-skill PE content. ALIGNED. Sync schedule remains disabled (containment). No PE prompt/branch/mutation; no Operational-Commons change; no shared-skills commit; no H/O resync.

Ship eligibility

NOT SHIP-ELIGIBLE until: (1) operator PASS sign-off on the re-review, (2) collision disposition for Hermes Management v1.0.0, (3) ship-time decisions (local-cli mapping, publish scope) + T11-T13 execution. Package evidence is otherwise complete.

Implementation evidence - narrow native stage-routing correction (PR #14, 2026-08-10)

Routing ambiguity defined: Core route is conceptual - it names the target stage but does not activate it. Observed native semantics: routed skills do not auto-load; explicit native load/activation is required; loading replaces the active stage skill. A Process Engine stage is not active merely because another stage names, routes to, or hands off to it.

PR: #14
Head: 5aad299fdb7628a4361abe159b7d14c5d12ac3a5 (branch fix/native-stage-routing from base 076dc4d…)
Exact scope: 5 files, +59/−7 - process-engine-core (routing-decision vs native-activation invariant), process-engine-pattern-author (handoff activates review), process-engine-review (transition-activation + REVISE loop activates pattern/review/trial), process-engine-trial (FAIL activates pattern/review; PASS retains operator Ship gate then activates ship), process-engine-triage (routed fixes/reviews/trials activate target stage). process-engine-ship unchanged (terminal). No references/personas/templates/governance; Summary Gate unchanged; Review/Trial verdict semantics unchanged; no runtime code/state machine/custom router.

Regressions (bounded prompt-level fixtures, labeled honestly): R1-R9 all satisfy the invariant (Core→Pattern activation; Pattern→Review activation; REVISE→Pattern→Review re-activation; PASS→Trial activation after operator sign-off; INCOMPLETE stays in Trial without Ship; FAIL activates correct revision stage; PASS without Ship GO does not deploy; activation failure stops with missing-activation evidence; already-active target not reloaded for ceremony). Hermes-H historical fixture: the emulated Review/Trial path would not be permitted under the corrected contract - Review/Trial could not have been executed from Core alone; loaded Review/Trial would have governed.

Structural validation: python3 tools/validate.py --repo . → STRUCTURAL VALIDATION PASS (8/8, exit 0).

Not done in this gate: no Hermes package Ship; Hermes Trial remains INCOMPLETE because T11-T13 (local-cli) are pending the operator's deploy-time local-cli mapping (separate decision); legacy Hermes Management skill untouched; no Operational-Commons change; sync schedule stays disabled. The Summary Gate finding remains behavioral adherence only (canonical Core was active and said "proceed only on confirmation"; the model attempted writes before GO) - this PR does not alter it.

Issue #13 remains open.

PR #14 correction - independent review `4898611629` addressed (2026-08-10)

Review finding: the initial PR #14 wording used unconditional activate/load, contradicting the stated R9 invariant (already-active target → no ceremonial reload).

Correction applied on new PR head 413db8bc5a1bc0672bd3568a59db9b3766f4ae81: the invariant is now ensure-active / load-if-needed - the exact canonical target stage must be proven natively active before execution; if not already proven active, natively activate/load it; if already proven active, do not reload for ceremony. "Already active" is precisely defined (exact canonical target proven to govern the current execution context; never memory, prose, another PE stage, stale version, prior workstream). Applied consistently across the existing 5-file scope (Core, Pattern Author, Review, Trial, Triage); Ship unchanged.

Direct R1-R9 regression receipts (prompt-level fixtures, honestly labeled; includes explicit R8 activation-failure and R9 + R9-near-miss evidence): #14 (comment)

Structural validation on new head: STRUCTURAL VALIDATION PASS (8/8, exit 0); GitHub workflow 31406808210 on exact head.

Hermes status unchanged (no Hermes work in this gate):

Hermes Maintenance fresh Review: PASS (revision 3).
Hermes Trial: INCOMPLETE - T11-T13 (local-cli surface) pending the operator's deploy-time local-cli mapping.
Package: NOT Ship-eligible until operator PASS sign-off + collision disposition + local-cli mapping + T11-T13 execution.

T11-T13 / local-cli remain held. No live PE reconciliation (live remains on canonical 076dc4d). Issue #13 remains open.

PR #14 merged + controlled live reconciliation + native stage-routing smoke (2026-08-10)

PR #14 merged at exact reviewed head 413db8bc5a1bc0672bd3568a59db9b3766f4ae81 (independent review 4898838490 PASS; exact-head workflow 31406808210 SUCCESS). Merge SHA 3a4d18ea7ace18eb310d2fc8b9e2a0ff946d015c; new main = 3a4d18ea…; parents 076dc4d… + 413db8bc…. PR-body evidence cleanup applied pre-merge (stale invariant corrected to ensure-active/load-if-needed; R1-R9 receipts labeled prompt-level contract fixtures, not runtime behavioral proof).

Controlled live reconciliation scope: exactly Core / Pattern Author / Review / Trial / Triage reconciled from new main; Ship, generator persona, dev persona, references/templates, policies/judge all unchanged. Rollback captured at sync-rollback-2026-08-10-pr14-reconcile/. Persisted read-back: all six live skills MATCH new main 3a4d18e… (canonical/live ALIGNED).

Native stage-routing smoke S1-S8 (live persisted skills): PASS

S1 Core→Pattern: native Loaded skill 'process-engine-pattern-author' receipt before authoring.
S2 Pattern→Review: native Loaded skill 'process-engine-review'; demonstrated Review-only contract (native parse receipt step A + allowed-field comparison step B) not inferable from Pattern.
S3 Review REVISE→Pattern→Review: 3 native loads; REVISE F-1 (non-spec version:) → Pattern rewrite → Review re-review PASS; no cross-stage emulation.
S4 Review PASS→Trial: native Trial load; applicable categories determined before execution; Ship never touched.
S5 Trial INCOMPLETE: unrun required category → INCOMPLETE (4/4 executed ≠ overall PASS); remains Trial; Ship not activated.
S6 already-active target: 0 load calls - no ceremonial reload; applied-identity receipt chain is the proof.
S7 active-stage near-miss: Trial active ≠ Review → native Review load before review behavior; near-miss not counted as target-active.
S8 activation failure: load process-engine-nonexistent-stage → native rejection not found or disabled, stop/report, state unchanged; not UNRUN (safe bounded probe existed).

Evidence distinction (kept separate): (a) prompt/static R1-R9 fixtures (5242758097) = contract assertions; (b) prior causal/runtime = Hermes Issue #13 stage-bound experiment; (c) new live behavioral = S1-S8 this gate. Stage-routing live conclusion: exact canonical target proven natively active before execution; load only when needed; never emulate; stop/report on failure.

Smoke telemetry: start 16:38:19Z → end ~17:00Z (~22 min incl. operator approval cycles); model deepseek-v4-flash (bundled); context occupancy ~624k tokens (ratio 0.045); authoritative consumed tokens/cost UNAVAILABLE (no provider surface); native loads 6; avoided redundant loads 1 (S6); operator questions 0; failures 0.

Hermes package remains (unchanged): fresh Review PASS; Trial INCOMPLETE (T11-T13 local-cli pending operator local-cli mapping); NOT Ship-eligible; no Hermes deployment/collision mutation performed. T11-T13 / local-cli decision held. Sync schedule remains disabled; sync defect owned by Operational-Commons #6. Issue #13 remains open.

Hermes Maintenance trial checkpoint close - RUNTIME CAPABILITY LIMITATION (2026-08-11)

Classification: RUNTIME CAPABILITY LIMITATION - NOT A PROCESS ENGINE PROMPT DEFECT.

Final classification

Item	State
Process Engine canonical	`3a4d18ea7ace18eb310d2fc8b9e2a0ff946d015c`
Hermes Maintenance Review	PASS (final re-review 2026-08-11; local-cli boundary revision accepted)
Capability preflight	bounded local execution path UNAVAILABLE in current target runtime
T11	UNRUN
T12	UNRUN
T13	UNRUN
Overall Trial	INCOMPLETE
Ship eligibility	NOT SHIP-ELIGIBLE

Evidence: actual `local-cli` → runtime semantics (read-only preflight)

The only Turnstone-native execution surface for local-cli is the bash tool (generator persona tool_allowlist: null = default interactive set, which includes bash).
Runtime characteristics: no command allowlist (verified: no allowed_commands/cmd_allowlist in turnstone core); no cwd/path restriction; arbitrary shell composition (pipes, $(...), redirection, multi-line); entire host filesystem as user vincent; full user environment incl. credential-readable paths.
The only guards are a soft destructive-pattern blocklist (trivially bypassable); tool policies are global and tool-name-level only (no command-level scoping).
Conclusion: effectively GENERIC SHELL, not a command-bounded executor. Package prompt guardrails are prompt-level; they do not constrain the runtime capability.

Excessive authority discovered → preflight STOP

Mapping local-cli → generic bash would expose full vincent-user shell authority (filesystem, environment, credential-readable paths) - materially beyond the reviewed bounded deployment/support-plane contract. Per the mapping gate's own rule, mapping stopped before enabling. Using generic bash would test a different authority model, not the reviewed package.

What this does and does not prove

Proves: the Trial capability preflight and gating behavior - the loaded canonical Trial skill (post-PR14) identified the required capability, preflighted actual target-runtime availability, and correctly yielded INCOMPLETE when the bounded path is unavailable. This is the canonical Trial contract working as designed.
Does NOT prove: the generated package's T11-T13 behavior (T11-T13 are UNRUN). In particular, the mapping-preflight STOP does not prove the package's T12 refusal behavior - that remains unexercised package behavior, not evidence.

Reason this is not a PE prompt defect

Canonical Review intentionally checks internal capability-path consistency but does not claim runtime availability (that is Trial's job). Canonical Trial's capability preflight is responsible for proving target-runtime availability and correctly yields INCOMPLETE when unavailable. The blocker is target-runtime capability availability, not a prompt gap. No Process Engine prompt correction is indicated.

Package status

The package remains internally Review-valid (frontmatter contract, parse receipts, destination evidence, capability-path coherence, boundary revision all PASS). The blocker is purely runtime capability availability. A future genuinely bounded Turnstone/external execution capability may make T11-T13 resumable, but that is a separate platform-capability workstream and must not be invented inside Process Engine.

CURRENT STATUS (2026-08-11) - supersedes earlier PASS framing

Readers should note: the original issue body and early comments describe the superseded first-pass Trial PASS / Ship-eligible framing from the initial (emulated-stage) run. That framing was overturned by the evidence-integrity audit (5236945355), the canonical stage-bound re-review + Trial completion (5237965731), and this checkpoint. Current truth: Hermes Maintenance = Review PASS, Trial INCOMPLETE (T11-T13 UNRUN - runtime capability limitation), NOT Ship-eligible. Historical comments are preserved as history; this CURRENT STATUS governs.

Hermes Maintenance - closure evidence (2026-08-11)

Status: SHIP PASS - deployed, read-back, predecessor superseded.

Pipeline evidence (all recorded, none rewritten)

Historical trial (2026-08-10): overall INCOMPLETE (former local-cli T11-T13 never executed).
Native Review: PASS (multiple revisions: MCP-only scope, delegation semantics, R7 3-agent routing correction).
Targeted trial: T1-T14 PASS; activation/routing 37/37 PASS; D1-D5 delegation regression PASS; U1-U3 focused regression PASS.
Reviewed behavioral artifact hash: 133b5a50…; post-trial package incl. evidence records: f3673d25….

Deployment (Turnstone-native, read-back verified)

Persona hermes-manager (enabled; MCP+memory; tool_allowlist hermes-gateway + turnstone-api).
Skills: hermes-management + 6 children (hermes-access-connectivity, hermes-health-observability, hermes-delegation, hermes-troubleshooting, hermes-update-maintenance, hermes-verification) - all enabled, named activation.
Prompt policy hermes-manager-prompt-policy (priority 1, no tool_gate).
Judge rule hermes-manager-advisory (low / review, advisory).
Stored content verified against the reviewed package (only trailing-newline normalization).

Supersession (reversible)

Legacy Hermes Management v1.0.0 disabled (enabled=false), content intact and recoverable, hidden from enabled discovery. Re-enable = rollback.

Runtime truth

Hermes v0.20.0 healthy; Hermes Management MCP 25 tools on 6/6 Turnstone contexts; no circuit open.
Gateway unchanged by Ship.

Stack model encoded

Turnstone = department head + active team member; direct Hermes control = bounded 25-tool MCP; Hermes-native capabilities delegable by task fit; Turnstone selects among itself / Hermes / OpenClaw by intent, capability, task fit, availability, authority, risk/gates, evidence. Package activation does not force Hermes selection.

Engine lessons (candidate)

When an execution-authority model changes, sweep the full body (not just declarations) for stale classifications - found twice in successive revisions.
"Direct control surface ≠ full agent capability ≠ delegable stack capability" generalizes to future agent packages (next: OpenClaw).

Downstream trial: Hermes Maintenance package (2026-08-10) #13

Description