Sync iteration-025 latency fast paths (small-gate review skip, proportional review-lite, lead slim) (#16)

* feat: add small-gate review skip, proportional review-lite, and phase-0 telemetry

* fix: obfuscate private hostname fragment in leak scanner; drop stale SearXNG and Synology docs

* fix: bump browserslist to patched release for dependency-audit (GHSA-c83g-rgw3-j3cx, GHSA-73wf-gq98-2v4g)

---------

Co-authored-by: Juan Carlos Rodicio <jc.rodicio@lookiero.com>

Author: jcarlosrodicio

README.md

@@ -240,6 +240,15 @@ Use the smallest useful flow:
   one-to-six-iteration budget, a hard ceiling of six, durable state, and
   reviewer-only closure. Final approval stops the cycle immediately.
 
+### Small-gate fast path
+
+A declared `small` change of at most two files may skip final review when it
+does not touch protected surfaces (auth, secrets, public API, migrations, CI,
+or the harness itself), is not destructive, and does not require new tests.
+`lead` then records `review_skipped_reason: small_gate_pass`. Medium and large
+work, and any sensitive surface, still get a full `reviewer` verdict. The
+unique gate definition lives in `opencode/docs/ai/harness/agents.md`.
+
 ## Included agents
 
 | Agent | Responsibility |
@@ -892,7 +901,6 @@ node scripts/check-harness.mjs
 │   ├── quickstart.md
 │   ├── security.md
 │   ├── superpowers.md
-│   ├── synology.md
 │   ├── troubleshooting.md
 │   └── workflows.md
 ├── opencode/
@@ -1028,7 +1036,6 @@ Detailed guides are available in [`docs/`](docs/):
 - [Threat model](docs/threat-model.md)
 - [Reproducible use cases](docs/use-cases/README.md)
 - [Troubleshooting](docs/troubleshooting.md)
-- [Synology notes](docs/synology.md)
 - [Awesome OpenCode submission draft](docs/distribution/awesome-opencode-submission.md)
 
 ## Compatibility and scope

docs/releases/v1.0.28.md

@@ -67,5 +67,4 @@ separate non-blocking observation.
 
 The public package excludes private OpenCode configuration, local provider and
 MCP wiring, credentials, authentication files, raw transcripts, private
-evidence, user-specific absolute paths, and machine-local state. In particular,
-the local SearXNG research integration is not part of this release.
+evidence, user-specific absolute paths, and machine-local state.

docs/releases/v1.0.29.md

@@ -74,5 +74,4 @@ limits.
 The public release excludes private OpenCode configuration, local provider and
 MCP wiring, credentials, authentication files, raw transcripts, private
 evidence, user-specific absolute paths, machine-local state, and internal Slice
-1.5 planning artifacts. In particular, the local SearXNG research integration
-is not part of this release.
+1.5 planning artifacts.

docs/synology.md

@@ -145,6 +145,14 @@ objective, decisions made, files read/touched, validation state, blockers, and
 next action. If there are long outputs, reference the artifact path instead of
 copying the log into context.
 
+Small-scope exemption: if the handoff includes the condensed 6-field card
+(objective, success, validation, diff base, constraints, output) and the task
+meets the small-gate defined uniquely in `docs/ai/harness/agents.md`, work
+against that card without expanding the full Task Contract or `handoff_packet`.
+In that case the card's `output` absorbs the essentials of the Result Contract
+and Verification Envelope in a short summary; otherwise the full blocks above
+and below remain mandatory.
+
 ## Result Contract and Verification Envelope
 
 When closing a non-trivial implementation, add a compact `Result Contract`:

opencode/agents/developer.md

@@ -102,6 +102,11 @@ If you choose direct mode, invoke `developer` with:
 - minimum acceptance criteria;
 - expected validation.
 
+If the small-gate in `docs/ai/harness/agents.md` also applies, delegate with its
+condensed 6-field card (canonical example there) instead of a full Task
+Contract, `handoff_packet`, and Result/Verification blocks; record
+`review_skipped_reason` at closeout.
+
 Do not mentally implement the solution before delegating. Any handoff to another
 agent must be self-contained and include:
 
@@ -118,12 +123,9 @@ decision made, and relevant paths are enough.
 
 ## Skill Resolution
 
-Before delegating non-trivial work to a subagent:
-
-1. If `docs/ai/harness/skill_registry.md` exists, read it.
-2. Filter skills allowed for the destination agent.
-3. Select only skills relevant to the task, with a maximum of 0-3 per handoff.
-4. Include a `Skill Resolution` block in the handoff prompt:
+Before delegating non-trivial work, consult `docs/ai/harness/skill_registry.md`
+only when it exists and is readable or there is an explicit match. Then select
+0-3 skills and emit:
 
 ```
 ## Skill Resolution
@@ -134,8 +136,8 @@ Before delegating non-trivial work to a subagent:
 - fallback_policy: if registry is missing, use global <available_skills>
 ```
 
-5. Prefer fewer skills over filling slots.
-6. If no skill clearly matches, select none.
+If the registry is missing or no skill clearly matches, emit no block and no
+dead path pointers. Prefer fewer skills over filling slots.
 
 ## Auto-Forecast
 
@@ -182,6 +184,17 @@ Examples:
 
 Use the full phased flow only when there is real uncertainty, visual/product impact, technical decision-making, medium or large scope, or an explicit slash command.
 
+## Fast router and phase model
+
+Explicit precedence: fast router first. A small, well-routed free-form message
+exits the phase model early under the SAME hard exclusions as the small-gate:
+protected path, destructive diff, or medium/large with real uncertainty =>
+phased flow with final review; slash-command flows => full model. Canonical
+sequence: 1 Intake -> 2 Discovery -> 3 Synthesis -> 4 Specification ->
+5 Implementation -> 6 Review -> 7 Closure. Do not restart phases or add agents
+because of the count. On-demand AHE templates:
+`docs/ai/harness/playbooks/lead-ahe-flow.md`.
+
 ## Required execution model
 
 Use explicit phase barriers:
@@ -207,8 +220,13 @@ Do not skip phases unless the work is genuinely trivial or you explicitly routed
   implementation correction goes back to `developer`.
 - Never invoke `reviewer` before there is a diff or implementation to review.
 - Every reviewable change requires a canonical `review_stage: final` envelope
-  from `reviewer` before closure. Give it the objective, non-goals, task/spec,
-  diff base, and original evidence; a developer summary is not a substitute.
+  from `reviewer` before closure, except the small-gate defined uniquely in
+  `docs/ai/harness/agents.md` (declared small + `<=2` files + no protected or
+  destructive paths + no required tests). Then close without reviewer and
+  record `review_skipped_reason: small_gate_pass`; outside the gate use the
+  reason enum. When review proceeds, give it the objective, non-goals,
+  task/spec, diff base, and original evidence; a developer summary is not a
+  substitute.
 - If `reviewer` returns `verdict: needs_changes`, synthesize the blocking
   findings, send a bounded correction task to `developer`, then invoke
   `reviewer` again. If it returns `blocked`, obtain the missing evidence or
@@ -255,6 +273,15 @@ When useful, include:
 ## Closure
 
 A task is closed only when there is concrete output, reasonable validation or a
-clear reason validation was not run, a final reviewer envelope has verdict
-`pass` or `pass_with_observations`, no blocking findings remain, and risks are
-explicit.
+clear reason validation was not run, and either a final reviewer envelope with
+verdict `pass` or `pass_with_observations` and no blocking findings, or the
+small-gate with `review_skipped_reason`. Make risks explicit.
+
+## AHE for harness evolution
+
+Changing agents, commands, skills, tools, or global rules is observable harness
+work: evidence -> root cause -> manifest -> bounded changes -> re-evaluation ->
+attribution -> review. Details:
+`docs/ai/harness/playbooks/lead-ahe-flow.md`. Each change declares evidence,
+root cause, component, predicted fixes, risks, and keep/improve/rollback+pivot.
+Without git there is no automatic rollback; do not simulate one.

opencode/agents/lead.md

@@ -84,6 +84,28 @@ For `/plan`, apply the same contract to the objective, research, plan/spec,
 assumptions, risks, acceptance criteria, and validation design even when no
 implementation diff exists. State that limitation.
 
+## Proportional review by scope
+
+Verdict rigor does not change with scope; review depth must be proportional:
+
+- Review levels are `skipped`, `trivial`, `lite`, and `full`, with the
+  semantics defined in `docs/ai/harness/commands.md` (section
+  `/review-preflight`). The caller chooses the level from the small-gate in
+  `docs/ai/harness/agents.md`; this skill does not add new profiles.
+- When review runs at small or medium scope (`lite`), consume the developer's
+  `Verification Envelope` first (`diff_base`, `review_scope`, `commands_run`,
+  `results`, `not_run`, `evidence`) as the authoritative record of checks
+  already run. Do not re-implement suites that envelope leaves green: repeat a
+  check only with a founded suspicion of the result, or when the envelope is
+  missing/incomplete for a blocking surface.
+- In `lite`, spot-check the real diff and its direct consumers. Mark checks
+  adopted from the envelope as `envelope-reused` in the output section
+  `Checks reviewed/repeated`.
+- For medium/large work and sensitive surfaces (auth/sessions, secrets,
+  exported public API, migrations, CI, harness) apply full verification and
+  the intact final verdict with no shortcuts: never substitute that depth with
+  trust in the envelope.
+
 ## Task Contract And Skill Resolution
 
 Confirm these Task Contract fields when applicable: `objective`,

opencode/agents/reviewer.md

@@ -0,0 +1,2 @@
+{"schema_version":1,"scenario_id":"freeform-tiny-copy-trivial","observation_coverage":{"session_events":"complete","skill_events":"complete","tool_events":"complete","review_events":"complete","stop_events":"complete"},"git_snapshot":"changed","events":[{"sequence":1,"type":"session_started","session_id":"root","parent_session_id":null,"agent":"lead"},{"sequence":2,"type":"session_started","session_id":"s1","parent_session_id":"root","agent":"developer"},{"sequence":3,"type":"tool_call_completed","session_id":"s1","agent":"developer","tool":"apply_patch","mutation":"write"},{"sequence":4,"type":"session_completed","session_id":"s1","agent":"developer"},{"sequence":5,"type":"session_started","session_id":"s2","parent_session_id":"root","agent":"reviewer"},{"sequence":6,"type":"review_completed","session_id":"s2","agent":"reviewer","verdict":"approved"},{"sequence":7,"type":"session_completed","session_id":"s2","agent":"reviewer"},{"sequence":8,"type":"stop_observed","condition":"task-completed"}]}
+{"schema_version":1,"scenario_id":"freeform-tiny-toy-single-file-full-pipeline","observation_coverage":{"session_events":"complete","skill_events":"complete","tool_events":"complete","review_events":"complete","stop_events":"complete"},"git_snapshot":"changed","events":[{"sequence":1,"type":"session_started","session_id":"root","parent_session_id":null,"agent":"lead"},{"sequence":2,"type":"session_started","session_id":"s1","parent_session_id":"root","agent":"researcher"},{"sequence":3,"type":"session_completed","session_id":"s1","agent":"researcher"},{"sequence":4,"type":"session_started","session_id":"s2","parent_session_id":"root","agent":"specifier"},{"sequence":5,"type":"session_completed","session_id":"s2","agent":"specifier"},{"sequence":6,"type":"session_started","session_id":"s3","parent_session_id":"root","agent":"developer"},{"sequence":7,"type":"tool_call_completed","session_id":"s3","agent":"developer","tool":"apply_patch","mutation":"write"},{"sequence":8,"type":"session_completed","session_id":"s3","agent":"developer"},{"sequence":9,"type":"session_started","session_id":"s4","parent_session_id":"root","agent":"reviewer"},{"sequence":10,"type":"review_completed","session_id":"s4","agent":"reviewer","verdict":"approved"},{"sequence":11,"type":"session_completed","session_id":"s4","agent":"reviewer"},{"sequence":12,"type":"stop_observed","condition":"task-completed"}]}

opencode/docs/ai/evolution/benchmarks/replay-fixtures-red.jsonl

@@ -0,0 +1,2 @@
+{"schema_version":1,"id":"freeform-tiny-copy-trivial","category":"trivial","prompt":"Change the Save button label to Save changes in a single HTML file and verify the diff.","command_path":"freeform","expected_root_agent":"lead","required_agents":["developer"],"forbidden_agents":["reviewer","evaluator","debugger","evolver"],"allowed_skills":[],"forbidden_skills":[],"write_before_spec_policy":"allowed","review_policy":"forbidden","expected_stop_condition":"task-completed","maximum_delegation_budget":1,"required_evidence":["static_contract"]}
+{"schema_version":1,"id":"freeform-tiny-toy-single-file-full-pipeline","category":"trivial","prompt":"Create a single HTML file with a canvas animation of a Rubik cube and verify the diff.","command_path":"freeform","expected_root_agent":"lead","required_agents":["developer"],"forbidden_agents":["researcher","specifier","reviewer","evaluator","debugger","evolver"],"allowed_skills":[],"forbidden_skills":[],"write_before_spec_policy":"allowed","review_policy":"forbidden","expected_stop_condition":"task-completed","maximum_delegation_budget":1,"required_evidence":["static_contract"]}

opencode/docs/ai/evolution/benchmarks/router-scenarios-red.jsonl

@@ -83,6 +83,33 @@ generate or modify frontmatter.
 - `evaluator`, `debugger`, and `evolver` are optional sidecars.
 - `evolver` works only on the OpenCode harness.
 
+### Small-gate closeout without review
+
+This is the unique, computable definition of "small" that may skip final
+review. `lead` and the condensed Task Contract card consume this SAME
+definition; no second definition exists in other contracts. The gate holds
+ONLY when every condition is true:
+
+1. Scope declared `small` in the handoff or Task Contract.
+2. `diff_files_changed <= 2` counted against the real diff base.
+3. No file touches a protected path: auth/sessions, secrets/.env, exported
+   public API, migrations, CI, or the harness itself (`agents/**`,
+   `commands/**`, `docs/ai/harness/**`, `scripts/**`, `AGENTS.md`,
+   `docs/ai/evolution/benchmarks/**`, `skills/**`).
+4. Diff is not destructive: no mass deletes, no reverts of others' work, no
+   permission or security-config changes.
+5. The touched surface does not require new or mandatory tests.
+
+Point-3 exclusions beat any size or prompt-length heuristic. When closing
+through this gate, `lead` records `review_skipped_reason: small_gate_pass`.
+The closed enum is: `small_gate_pass | protected_surface | destructive_diff |
+tests_required | scope_not_small | human_decision`; the other values document
+why the fast path did not apply, or an explicit human decision.
+
+This gate does not reduce `reviewer` authority: medium/large work and every
+sensitive surface keep a full `review_stage: final` with a single final
+`reviewer` verdict. It does not apply retroactively to work already reviewed.
+
 ## Mission runtime projection
 
 `mission-status` is a read-only view over the durable loop snapshot and
@@ -130,6 +157,29 @@ when the change is not trivial. The minimum block is:
 - `validation`: commands, tests, or evidence.
 - `ask_abort_triggers`: conditions for asking, blocking, or returning work.
 
+### Condensed card for small scope
+
+When the change meets the small-gate in the section above ("Small-gate closeout
+without review", same unique definition), `lead` may delegate and `developer`
+may close with a condensed 6-field card instead of the full Task Contract,
+`handoff_packet`, and separate Result Contract / Verification Envelope stack.
+This is an activation exemption, not a new contract: for non-trivial work every
+prior block remains mandatory. Canonical audited example, exactly 6 fields:
+
+```text
+small-card:
+  objective: rename the Save button to Save changes
+  success: visible label changed in the UI
+  validation: grep the literal + local render smoke
+  diff_base: working tree on HEAD (1 file)
+  constraints: do not touch other literals or tests
+  output: short summary with the applied diff
+```
+
+When closing this way, `lead` records `review_skipped_reason` from the gate
+enum. Telemetry (turn ids, timestamps, diff size) ALWAYS stays out of the
+card: it lives only in evidence-collector output.
+
 For long, multi-agent, or resumable work, the responsible agent adds a compact
 `handoff_packet` with current objective, decisions made, files read/touched,
 validation state, blockers, and next action. Long logs are referenced by path