alexsoyes · Sep 30, 2026 · Sep 30, 2026 · Sep 30, 2026 · Oct 5, 2026 · Oct 5, 2026
@@ -8,29 +8,17 @@ argument-hint: conversation | export
 ```mermaid
 flowchart LR
   start([conversation ID or export]) --> read-conversation
-  read-conversation -->|complete| scopes{two or more scopes?}
+  read-conversation -->|complete| recommend --> target-edits
   read-conversation -->|unavailable| unavailable([stop])
-  scopes -->|no| recommend-local[recommend locally] --> target-edits
-  scopes -->|yes| isolation{isolated artifact context?}
-  isolation -->|yes| recommend-parallel[recommend in parallel] --> target-edits
-  isolation -->|no| recommend-local
   target-edits --> question([ask next intent]) --> stop([stop])
 ```
 
 ## Actions
 
-Run the flow above. Read only the next action file.
+Run all three actions in order without confirmation. Read only the next file in `actions/`. Keep the project read-only; write only the unique temporary report.
 
-| Action | Does |
+| Order | Action |
 | --- | --- |
-| read-conversation | freeze complete evidence and measure visible cost |
-| recommend | analyze relevant scopes and merge grounded findings |
-| target-edits | render minimal edits and an executable prompt |
-
-## Transversal rules
-
-- After resolving the source, run all three actions without pausing for confirmation.
-- Analyze only a complete conversation and the exact skills or documents it names.
-- Exclude every current or previous `improve` invocation and report from the analysis evidence.
-- Never invent time, tokens, source coverage, or document status.
-- Do not modify, stage, or persist any project file.
+| 1 | `01-read-conversation.md` |
+| 2 | `02-recommend.md` |
+| 3 | `03-target-edits.md` |
@@ -1,35 +1,23 @@
 # 01 - Read conversation
 
-Load complete conversation evidence once and profile its visible cost.
+Freeze complete evidence and measure visible cost.
 
 ## Input
 
 A current conversation, an exact conversation ID, or a complete transcript export.
 
 ## Output
 
-An evidence boundary, a `## Timing` table with `Activity | Observed time | Share | Evidence`, a `## Usage` table with `Metric | Value | Evidence`, and a scope index.
+An evidence boundary, a `## Timing` table with `Activity | Observed time | Share | Evidence`, a `## Usage` table with `Metric | Value | Evidence`, and a scope index of encountered resources.
 
 ## Process
 
-1. **Resolve.** Choose the host-specific transcript source from [conversation sources](../assets/conversation-sources.md).
-   - Stop when no complete transcript can be resolved for the exact conversation.
-2. **Bound.** Freeze the evidence at the message before the current invocation, or at the supplied export boundary.
-3. **Read.** Load every in-scope message, tool call, tool result, and timestamp from that transcript.
-4. **Index.** Record relevant turns and invoked or named skills and knowledge files for downstream analysis.
-5. **Measure.** Calculate visible time and usage only from timestamps, elapsed records, or host usage data.
-   - Group explicit tool activity as `research and diagnosis`, `implementation`, or `validation`; keep mixed or unknown time `unattributed`.
-   - Include tokens, requests, and monetary cost only when the source exposes them.
-   - Never label unattributed wait as reasoning time.
-6. **Render.** Mark a missing metric `unavailable` and order known activity times from longest to shortest.
-
-## Test
-
-| Case | Pass |
-| --- | --- |
-| A conversation is analyzed | its source resolves to one exact complete transcript |
-| The same conversation is analyzed again | its evidence excludes every `improve` invocation and report |
-| A timing metric is shown | its evidence identifies visible timestamp or elapsed records |
-| A metric is unavailable | the report does not estimate or call it reasoning time |
-| A usage metric is shown | its evidence identifies the host record that exposes it |
-| A scope is indexed | it identifies exact turns or artifact paths, not a generated summary |
+1. **Resolve.** Use the host route in [conversation sources](../assets/conversation-sources.md). Prefer its complete export or host reader. Match the exact session ID before direct storage; search only for that ID, never unrelated sessions, configuration, or authentication. Stop unless the exact complete transcript is available.
+2. **Freeze.** End before this invocation, or at the export boundary. Exclude every `improve` invocation and report.
+3. **Read once.** Load every in-scope message, tool call, result, and timestamp. Index relevant turns; invoked skills and their used actions or resources; applicable `AGENTS.md`; and task-relevant memory references. Do not read maintained sources yet.
+4. **Measure.** Use only timestamps, elapsed records, and exposed host usage.
+   - Group tool time as `research and diagnosis`, `implementation`, `validation`, or `unattributed`.
+   - Include tokens, requests, and cost only when exposed; otherwise mark them `unavailable`.
+   - Separate background-process lifetime from blocking time.
+   - Treat private reasoning duration as unavailable unless exposed directly; never infer it.
+5. **Deliver.** Order known activity times descending and cite every value.
@@ -1,50 +1,37 @@
 # 02 - Recommend
 
-Analyze each relevant scope with the smallest grounded change.
+Reduce avoidable work without reducing reliability.
 
 ## Input
 
-The evidence boundary, timing and usage tables, complete conversation, and scope index.
+The evidence boundary, timing and usage tables, complete conversation, and resource index.
 
 ## Output
 
-A `## Recommendations` table with `ID | Question | Type | Diagnostic | Evidence | Smallest change | Target | Saving`.
+An updated `Scope | Status | Evidence` index, a `## Recommendations` table with `ID | Focus | Type | Diagnostic | Evidence | Smallest change | Source | Saving`, and non-actionable limitations.
 
 ## Process
 
-1. **Scope.** Select `behavior`, plus `skill` and `knowledge` only when the scope index names relevant artifacts.
-2. **Dispatch.** Analyze locally when one scope exists; otherwise dispatch one read-only analyst per scope in parallel.
-   - Prefer a lightweight available model and low reasoning effort when the host supports per-agent overrides; otherwise inherit the run defaults.
-   - Give the behavior analyst the complete frozen transcript; give artifact analysts the same boundary, indexed turns, and exact artifact paths.
-   - Request isolated or minimal context for artifact analysts when supported; otherwise analyze every scope locally instead of duplicating the transcript.
-   - Do not dispatch an analyst with no relevant evidence.
-   - Require table rows only and forbid file writes.
-3. **Question.** Make each analyst answer every prompt for its scope.
+1. **Verify.** Resolve and read indexed maintained skill resources, applicable `AGENTS.md`, and task-relevant memory references once; reuse a still-valid read. Add and read only relevant paths from the project memory index; never scan unrelated memory. Neither judge from nor target caches or installs. Account for host transforms. Mark only read sources `checked`; mark unresolved or skipped ones `missing` or `not reviewed`.
+2. **Scope.** Always assess `behavior`; add `skill` or `knowledge` only for checked indexed sources.
+3. **Dispatch.** With multiple isolated scopes, run one read-only analyst per scope in parallel. Otherwise analyze locally.
+   - Prefer a cheap model and low reasoning effort when per-agent overrides are supported; otherwise inherit host defaults.
+   - Give the behavior analyst the frozen transcript. Give artifact analysts the boundary, indexed turns, and verified excerpts with evidence; allow a targeted read-only refresh only when validity is uncertain.
+   - Give every analyst steps 4–6 and require table rows.
+4. **Reflect.** Ask yourself, for each scope; keep this assessment internal:
    - How could the next run be faster or better?
-   - What information should be removed or clarified?
-   - Where should the change live?
-   - How could it save time or tokens?
-   - What work was counterproductive?
-4. **Verify.** Read a named skill or knowledge file before assessing its information.
-5. **Assess.** Label relevant information `obsolete`, `over-specific-or-time-bound`, `duplicate`, `inconsistent`, `counterproductive`, or `correct`.
-   - Use `correct` when no evidence supports another label, and never render it as a recommendation.
-6. **Merge.** Deduplicate findings across scopes and verify only their cited evidence against the frozen source.
-7. **Render.** Order by question then `behavior`, `skill`, `knowledge`, and describe each change with the fewest unambiguous words.
-   - Use `skill`, `behavior`, `knowledge`, or `tooling` as the target type.
-   - State `time`, `tokens`, `both`, or `unknown` as its saving.
-   - Render `no change` when a scope has no evidence-backed recommendation.
-
-## Test
-
-| Case | Pass |
-| --- | --- |
-| A question is shown | it is answered from conversation evidence |
-| One relevant scope exists | no parallel analyst is dispatched |
-| Several relevant scopes exist | their analysts run in parallel and return the same columns |
-| An artifact analyst cannot receive isolated context | every scope is analyzed locally instead of duplicating the transcript |
-| A type is shown | it is `skill`, `behavior`, `knowledge`, or `tooling` |
-| Information is assessed | it has one allowed label backed by evidence |
-| Information is correct | its scope says `no change` and no recommendation is rendered |
-| A named target is assessed | the target file was read before the verdict |
-| A saving is shown | it is categorical and never an invented amount |
-| The same evidence is analyzed again | recommendations keep the same order and do not cite an earlier `improve` report |
+   - What should be removed or clarified? What was counterproductive?
+   - Where should change live? How could it save time or tokens?
+   - Which back-and-forth, bottlenecks, or tool calls could be removed, batched, parallelized, or replaced?
+5. **Assess.** Use `obsolete`, `over-specific-or-time-bound`, `duplicate`, `inconsistent`, `counterproductive`, or `correct`.
+   - Compare observed work with the shortest reliable path to the requested result. Keep checks for mutable state or unresolved uncertainty; reuse still-valid results instead of repeating research.
+   - Raw call count never proves waste. Ground the smallest general correction in evidence.
+   - Recommend memory only for concise, durable knowledge that prevents recurring rediscovery, never transient state or one-off implementation details.
+   - For skills, especially knowledge: delete evidence-backed waste or inconsistency first, without quota; then consolidate or clarify.
+   - Reuse or strengthen existing content; add only for a demonstrated gap it cannot cover. Preserve useful context and requirements; classify sufficient content as `correct` => `no change`.
+   - Claim time savings from blocking-time or resource-impact evidence, never background-process uptime alone.
+6. **Deliver.** Deduplicate, verify cited evidence, and order by focus, then `behavior`, `skill`, `knowledge`. Report findings, not the questionnaire, in the fewest actionable words.
+   - Every recommendation needs an exact maintained source path, current excerpt, minimal correction or explicit `+`/`−` edit, and a short target/action prompt.
+   - If no source resolves or no useful persistent edit exists, report a non-actionable limitation.
+   - Type: `skill`, `behavior`, `knowledge`, or `tooling`.
+   - Saving: `time`, `tokens`, `both`, or `unknown`.
@@ -1,35 +1,20 @@
 # 03 - Target edits
 
-Map recommendations to minimal edits and render the report.
+Map minimal edits and render the report.
 
 ## Input
 
-The timing, usage, and recommendations tables.
+The timing, usage, recommendations, and scope index tables.
 
 ## Output
 
-An HTML report in a temporary directory with local `report.css` and `report.js`, plus its path and the next-intent question.
+An HTML report in a temporary directory with local `report.css` and `report.js`, plus its path and a next-intent request.
 
 ## Process
 
-1. **Target.** Map each file-targeted recommendation to a real project path.
-   - Omit behavior-only recommendations from this table.
-2. **Summarize.** Build a `Fichier | + Ajout | − Retrait ou clarification` table.
-   - Consolidate repeated targets and render each diagnostic's smallest change.
-   - Render one `Aucun fichier recommandé | — | —` row when no recommendation needs a file edit.
-3. **Render.** Fill [the report template](../assets/report-template.html) with only measured values and grounded findings, then copy its local CSS and JavaScript beside it.
-   - Remove every sample value and sample finding from the produced report.
-   - HTML-escape every injected value; allow only template-owned markup and local asset references.
-   - Write only to a unique temporary directory, never the project.
-4. **Return.** Provide the report path and end with this exact question: `Quel changement d’intention général, même minime, appliquons-nous au prochain run pour rendre notre amélioration cumulative et mesurable ?`
-
-## Test
-
-| Case | Pass |
-| --- | --- |
-| A file is shown | it is the target of an earlier recommendation |
-| No file needs editing | the empty-table row is shown |
-| The report is rendered | its HTML, CSS, and JavaScript resolve locally with no network dependency |
-| Conversation evidence is rendered | it is escaped as text and cannot create markup, scripts, or URLs |
-| A value is unavailable | the report says `unavailable` instead of showing sample data |
-| The run closes | the returned chat response ends with the exact next-intent question |
+1. **Target.** Consolidate the provided recommendations by source into `File | + Added | − Removed or clarified`; use `No recommended file | — | —` when empty. Never resolve, judge, or apply sources here.
+2. **Render.** Fill [the report template](../assets/report-template.html) from the provided data; remove its sample and copy its local CSS and JavaScript into the allowed unique temporary directory.
+   - Escape injected values; allow only template markup and local assets. Localize the HTML `lang`, visible text, and dynamic `data-label-*` text to the user's language.
+   - Put findings first with diagnostic/type tags, filters, accept controls, the target table, and an editable global prompt. Copy each recommendation's short target/action prompt into `data-prompt` without redrafting it.
+   - Show source paths and minimal edits. Collapse raw evidence, secondary metrics, timing, coverage, source records, and limitations; keep the run ID in source records and avoid repeated metadata.
+3. **Return.** Return its path. Ask the user, in their language, which small general change in intent to apply next run for cumulative, measurable improvement.
@@ -7,11 +7,3 @@ Use this static routing map to load one complete conversation. Do not fill or co
 | Codex | host thread reader with the exact thread ID, then TUI `/export` Markdown | `CODEX_HOME/history.jsonl`, normally `~/.codex/history.jsonl` | TUI `/status` estimated thread credits or cost; tool elapsed records when exposed |
 | Claude Code | `/export` text, or the documented script interface for an exact session ID | `~/.claude/projects/{project}/{session-id}.jsonl` | transcript timestamps and tool records when exposed |
 | OpenCode | `opencode export {session-id} --sanitize` JSON | `~/.local/share/opencode/project/{project-slug}/storage/` | `opencode stats`; `~/.local/share/opencode/log/` |
-
-## Rules
-
-- Prefer the complete export or host reader over direct storage parsing.
-- Use direct storage only after matching one exact session ID.
-- Search shared history or logs only for the exact session ID and ignore unrelated records.
-- Never enumerate unrelated sessions or inspect configuration or authentication files.
-- Treat private reasoning duration as unavailable unless the host exposes it directly.