@event4u/agent-config 9.9.0 → 9.11.0

package/dist/agent-src/commands/agents/user/review.md

@@ -7,7 +7,7 @@ visibility: internal
 cluster: agents
 sub: user
 skills: [agents]
-description: List buffered observations from .agent-user.observations.jsonl with numbered options to inspect or accept individually.
+description: List buffered observations from the project-local and global observation buffers with numbered options to inspect or accept individually.
 suggestion:
   eligible: false
   rationale: "Cluster sub-command — reached via its cluster head's routing or its explicit /cluster:sub name; not independently suggested (surface-consolidation)."
@@ -19,14 +19,17 @@ packs:
 
 # /agents user review
 
-List the buffered observations in `.agent-user.observations.jsonl`
-and let the user choose which to inspect or accept.
+List the buffered observations across BOTH observation buffers —
+project-local `.agent-user.observations.jsonl` and the global
+`~/.event4u/agent-config/user/observations.jsonl` (ADR-138,
+road-to-global-user-memory Phase 2) — and let the user choose which to
+inspect or accept.
 
 Use when:
 
 - You want to see what the agent has learned about your preferences.
-- You suspect `.agent-user.md` is out of date and want a curated
-  diff before editing.
+- You suspect `.agent-user.md` (or the global `profile.md`) is out of
+  date and want a curated diff before editing.
 - You want to dismiss observations the agent collected.
 
 Read-only by itself — actual changes go through
@@ -35,57 +38,90 @@ Read-only by itself — actual changes go through
 
 ## Steps
 
-### 1. Locate buffer
+### 1. Locate buffers
 
 ```bash
 ls .agent-user.observations.jsonl 2>/dev/null
 ```
 
-| State | Action |
+Resolve the global buffer's path via `user_global_observations.
+resolveGlobalObservationsPath()` — it honours the config-home override
+env var and the legacy fallback root exactly like every other global
+artefact, so never hard-code either path here.
+
+| State (either buffer) | Action |
 |---|---|
-| Missing or empty | Print "No buffered observations. The agent has not learned anything new." and stop |
-| Present | Proceed |
+| Both missing or empty | Print "No buffered observations. The agent has not learned anything new." and stop |
+| At least one present | Proceed with whichever buffer(s) exist |
 
 ### 2. Parse + group
 
-Read every line as JSON. Drop malformed lines silently (one-line
-warning at the end with the count). Group by `field`:
+Read every line as JSON from BOTH buffers via
+[`user_global_observations.readGlobalObservations`](../../../../src/scripts/_lib/user_global_observations.ts)
+for the global one, and the equivalent tolerant parse for the
+project-local file. Drop malformed lines silently (one-line warning at
+the end with the combined count; `readGlobalObservations` already
+reports `droppedMalformed` / `droppedUnknownField`). Group by `field`
+ACROSS both buffers, tagging each observation with its layer:
 
 ```
-.agent-user.observations.jsonl — {n} observations across {k} fields
+Buffered observations — {n} across {k} fields ({p} project-local, {g} global)
 
   1. style.pace          ({n}× since {oldest_ts})
-     latest suggest: "rapid" — evidence: user said 'mach kürzer' 3× this session
+     latest suggest: "rapid" [global] — evidence: user said 'mach kürzer' 3× this session
   2. identity.nickname   ({n}× since {oldest_ts})
-     latest suggest: "Matze" — evidence: user signed last 3 messages "— Matze"
+     latest suggest: "Matze" [project] — evidence: user signed last 3 messages "— Matze"
   3. language            ({n}× since {oldest_ts})
-     latest suggest: "de" — evidence: last 12 messages in German, .agent-user.md says "en"
+     latest suggest: "de" [global] — evidence: last 12 messages in German, profile says "en"
+```
+
+Sort by frequency (most observations first, combined across both
+layers), then by recency. When the SAME field has observations in both
+layers, list the most-recent one first and note both counts.
+
+### 2a. Promotion candidates (road-to-global-user-memory Phase 3)
+
+Separately from the field grouping above, call
+[`findPromotionCandidates`](../../../../../src/scripts/_lib/user_global_observations.ts)
+over the global buffer's entries. Any observation with `seen_count ≥ 3`
+(`PROMOTION_SEEN_COUNT_THRESHOLD`) is project-attributed, human-confirmed
+recurring evidence across that many DIFFERENT projects — surface it as
+its own block, ahead of the field list:
+
+```
+Promotion candidates — seen in 3+ projects:
+
+  1. "always use pnpm instead of npm for installs" — seen in 3 projects (acme-web, acme-api, acme-mobile)
 ```
 
-Sort by frequency (most observations first), then by recency.
+This is candidacy only, never a promotion — crossing the threshold makes
+`/agents user accept` able to offer promoting it, nothing more (see
+[`agent-user-schema.md § Project attribution`](../../../../../docs/contracts/agent-user-schema.md#project-attribution-road-to-global-user-memory-phase-3)).
 
 ### 3. Ask
 
 ```
-> 1. Accept the most-frequent suggestion ({field} → {value})
+> 1. Accept the most-frequent suggestion ({field} [{layer}] → {value})
 > 2. Inspect a specific field
-> 3. Clear the buffer (discard all)
-> 4. Cancel
+> 3. Clear a buffer (project-local, global, or both)
+> 4. Promote a candidate from step 2a (only shown when ≥ 1 exists)
+> 5. Cancel
 ```
 
 | Choice | Action |
 |---|---|
-| 1 | Hand off to [`/agents user accept`](accept.md) with the chosen field |
-| 2 | Print every observation for that field with evidence + ts, then re-ask |
-| 3 | Truncate `.agent-user.observations.jsonl` to zero bytes; print confirmation |
-| 4 | Stop without changes |
+| 1 | Hand off to [`/agents user accept`](accept.md) with the chosen field AND layer |
+| 2 | Print every observation for that field (both layers) with evidence + ts, then re-ask |
+| 3 | Ask which buffer(s) to truncate, then truncate the chosen file(s) to zero bytes; print confirmation |
+| 4 | Hand off to [`/agents user accept`](accept.md) with the chosen promotion candidate — `accept` collects the mandatory `promotion_reason` before writing anything |
+| 5 | Stop without changes |
 
 One question per turn. Wait for the user's number.
 
 ### 4. Privacy-floor verify
 
 Before printing any observation's `evidence` text, scan it for the
-[exclusions list](../../../../../../../../docs/contracts/agent-user-schema.md#explicit-exclusions).
+[exclusions list](../../../../../docs/contracts/agent-user-schema.md#explicit-exclusions).
 Match → replace the offending substring with `[redacted]` in the
 rendered output.
 
@@ -100,14 +136,15 @@ turn.
 
 ## Rules
 
-- Read-only on `.agent-user.md`. Only `accept` writes to the persona
-  file.
-- Buffer is the only mutable artefact (truncate via option 3).
+- Read-only on `.agent-user.md` AND the global `profile.md`. Only
+  `accept` writes to either persona file.
+- Both buffers are the only mutable artefacts (truncate via option 3).
 - Mirror the user's language for prompts per
   [`language-and-tone`](../../../../dist/agent-src/rules/language-and-tone.md).
 
 ## See also
 
-- Schema + buffer contract: [`agent-user-schema § Observation buffer`](../../../../../../../../docs/contracts/agent-user-schema.md#observation-buffer).
+- Schema + buffer contract: [`agent-user-schema § Observation buffer`](../../../../../docs/contracts/agent-user-schema.md#observation-buffer).
+- Global buffer implementation: [`user_global_observations.ts`](../../../../src/scripts/_lib/user_global_observations.ts).
 - Parent: [`/agents user`](../user.md).
 - Sibling: [`/agents user accept`](accept.md), [`/agents user update`](update.md).

package/dist/agent-src/commands/agents/user/show.md

@@ -7,7 +7,8 @@ visibility: internal
 cluster: agents
 sub: user
 skills: [agents]
-description: Read-only render of .agent-user.md — prints the persona summary the host agent loads at session start.
+description: Read-only render of the effective (merged) user profile — global profile.md plus project .agent-user.md. --audit renders the global layer raw for delete/revoke decisions.
+argument-hint: "[--audit]"
 suggestion:
   eligible: false
   rationale: "Cluster sub-command — reached via its cluster head's routing or its explicit /cluster:sub name; not independently suggested (surface-consolidation)."
@@ -19,50 +20,71 @@ packs:
 
 # /agents user show
 
-Read-only render of the project-root `.agent-user.md` per
-[`docs/contracts/agent-user-schema.md`](../../../../../../../../../../../docs/contracts/agent-user-schema.md).
+Read-only render of the **effective, merged** user profile — the
+global `~/.event4u/agent-config/user/profile.md` layer plus the
+project-root `.agent-user.md` layer — per
+[`docs/contracts/agent-user-schema.md`](../../../../../docs/contracts/agent-user-schema.md)
+and its Loader contract / merge rule (ADR-138).
 
 Use when:
 
 - You want to see what persona the host agent currently loads.
-- You want to confirm `last_updated` is fresh (≤90 days).
+- You want to confirm the **effective** `last_updated` is fresh (≤90 days).
 - You want a paste-ready summary for handoff or onboarding.
 
 Does **not** edit, observe, or buffer anything. Pure read.
 
+Called with `--audit` (road-to-global-user-memory Phase 4): renders the
+GLOBAL layer's raw holdings instead of the merged effective profile —
+see [§ Audit mode](#audit-mode---audit) below. Use `--audit` when
+deciding what to delete via [`/agents user delete`](delete.md).
+
 ## Steps
 
-### 1. Locate the file
+### 1. Locate both layers
 
 ```bash
 ls .agent-user.md 2>/dev/null
+ls ~/.event4u/agent-config/user/profile.md 2>/dev/null
+ls ~/.config/agent-config/user/profile.md 2>/dev/null   # legacy fallback
 ```
 
-| State | Action |
-|---|---|
-| Present | Proceed |
-| Missing | Print "No `.agent-user.md` found at project root. Run `/agents user init` to create one." and stop |
+| Project file | Global file | Action |
+|---|---|---|
+| Present | Present or absent | Proceed — project layer wins per field it declares |
+| Missing | Present | Proceed — render sourced entirely from the global layer |
+| Missing | Missing | Print "No `.agent-user.md` and no global profile found. Run `/agents user init` to create one, or set up the global profile." and stop |
 
-### 2. Parse frontmatter
+### 2. Parse frontmatter — per layer
 
-Parse the YAML frontmatter and the body (everything after the second
-`---`). Validate against the locked v1 schema:
+Parse each present layer's YAML frontmatter and body (everything after
+the second `---`) independently. Per layer:
 
 - `version` is `1`.
-- `identity.name`, `language`, `role`, `style.pace`,
-  `voice_sample`, `last_updated` are all present.
-- File is ≤100 lines total.
+- File is ≤100 lines total — the cap applies **per layer**, never as a
+  shared total across both files.
+
+Any violation → print a one-line warning identifying the layer, the
+missing/malformed field, and continue with the render (so the user can
+fix it via `/agents user update`).
+
+### 3. Merge
 
-Any violation → print a one-line warning identifying the missing /
-malformed field and continue with the render (so the user can fix it
-via `/agents user update`).
+Apply the deepest-wins merge rule per
+[`agent-user-schema.md § Global profile layer`](../../../../../docs/contracts/agent-user-schema.md#global-profile-layer-adr-138):
+for each field (`identity.name`, `language`, `role`, `style.pace`,
+`voice_sample`, `last_updated`), the project layer's value wins when
+present, otherwise the global layer's value is used. `# Notes`
+concatenates both layers' text under `[global]` / `[project]` markers
+when both carry content.
 
-### 3. Render
+### 4. Render
 
-Print the persona in this exact shape:
+Print the merged persona in this exact shape:
 
 ```
-.agent-user.md  ({n} lines, last_updated: YYYY-MM-DD{staleness_marker})
+Effective user profile  ({staleness_marker})
+  sources: identity.name={global|project}, style.pace={global|project}, …
 
   Identity   : {nickname or name}  ({name} if nickname is set)
   Language   : {language}
@@ -75,15 +97,20 @@ Print the persona in this exact shape:
 
   Notes
   ─────────────
-  {body, indented 2 spaces; "(empty)" if no notes}
+  {merged notes, indented 2 spaces; "(empty)" if neither layer has notes}
 ```
 
-Where `{staleness_marker}` is:
+Where `{staleness_marker}` reflects the **effective** `last_updated`
+(the merged value, not necessarily the project file's own date):
 
-- empty when `last_updated` is within 90 days.
+- empty when within 90 days.
 - ` ⚠️  >90 days` when older (per the schema staleness rule).
 
-### 4. Loader hint
+The `sources:` line lists only the fields that resolved to a
+non-default value, so a project-only or global-only user sees a
+single-source line and is not confused by an empty cascade.
+
+### 5. Loader hint
 
 If the host-agent loader has NOT yet picked up the file this session
 (detect via session memory if available), print one line:
@@ -94,13 +121,39 @@ If the host-agent loader has NOT yet picked up the file this session
 
 Otherwise omit — agent already knows.
 
-### 5. Stop
+### 6. Stop
 
 Do NOT chain to other `/agents user *` commands. Do NOT commit.
 
+## Audit mode (`--audit`)
+
+Road-to-global-user-memory Phase 4. Renders what the GLOBAL layer
+currently holds — `profile.md`'s fields, the observation buffer's entry
+count and per-field counts, promotion candidates, and the revocation
+ledger's tombstone count — without the user reading raw JSONL. This is
+the global layer only; the project-local `.agent-user.md` never leaves
+the project it lives in, so there is nothing to audit for it here.
+
+1. Call
+   [`renderGlobalMemoryAudit()`](../../../../../src/scripts/_lib/user_global_memory_audit.ts).
+2. Print its `text` verbatim — every value it contains already passed
+   the same
+   [`knowledge_global_redaction.redaction_scan`](../../../../../src/scripts/_lib/knowledge_global_redaction.ts)
+   gate the write path runs; nothing here needs a second scan.
+3. If `revocationCount > 0`, add one line: "Run `/agents user delete
+   --show-revoked` to see what was deleted and why." (a stub for a
+   future `--show-revoked` flag; today the count is sufficient — the
+   full ledger is inspectable via
+   [`user_global_revocations.loadTombstones`](../../../../../src/scripts/_lib/user_global_revocations.ts)
+   for a maintainer who needs it).
+4. Stop. Do NOT chain to `--audit`'s own delete follow-up — hand the
+   user to [`/agents user delete`](delete.md) explicitly if they name
+   something to remove.
+
 ## Rules
 
-- Read-only. Never write `.agent-user.md` from this command.
+- Read-only. Never write `.agent-user.md` or the global `profile.md`
+  from this command — including `--audit` mode.
 - Never print fields the schema does not define — even if they exist
   in the file. (Forward-compat: an unexpected field is a warning, not
   a render target.)
@@ -110,6 +163,6 @@ Do NOT chain to other `/agents user *` commands. Do NOT commit.
 
 ## See also
 
-- Schema: [`agent-user-schema`](../../../../../../../../../../../docs/contracts/agent-user-schema.md).
+- Schema: [`agent-user-schema`](../../../../../docs/contracts/agent-user-schema.md) — see § Global profile layer for the merge rule this command renders, and § Delete, revoke, and audit for `--audit`'s contract.
 - Parent: [`/agents user`](../user.md).
-- Sibling: [`/agents user init`](init.md), [`/agents user update`](update.md).
+- Sibling: [`/agents user init`](init.md), [`/agents user update`](update.md), [`/agents user delete`](delete.md) — the delete/revoke counterpart `--audit` informs.

package/dist/agent-src/commands/agents/user/update.md

@@ -61,7 +61,7 @@ When called with `--validate` (or after the user confirms "done"):
 
 1. Read `.agent-user.md`.
 2. Parse frontmatter; check every required field per
-   [`agent-user-schema § Field reference`](../../../../../../../docs/contracts/agent-user-schema.md#field-reference).
+   [`agent-user-schema § Field reference`](../../../../../docs/contracts/agent-user-schema.md#field-reference).
 3. Check file size ≤100 lines.
 4. Run the privacy-floor scan (same as `init` step 4): credentials,
    third-party PII, financial figures, health/legal status.
@@ -99,6 +99,6 @@ Do NOT commit. Do NOT chain to another `/agents user *` command.
 
 ## See also
 
-- Schema: [`agent-user-schema`](../../../../../../../docs/contracts/agent-user-schema.md).
+- Schema: [`agent-user-schema`](../../../../../docs/contracts/agent-user-schema.md).
 - Parent: [`/agents user`](../user.md).
 - Sibling: [`/agents user init`](init.md), [`/agents user show`](show.md), [`/agents user review`](review.md).

package/dist/agent-src/commands/agents/user.md

@@ -25,6 +25,18 @@ Sub-dispatcher for the user-persona file
 single, project-root, gitignored Markdown file that captures who the
 user is and how they want the agent to address them.
 
+Since ADR-138, a weaker global layer
+(`~/.event4u/agent-config/user/profile.md`) sits beneath the
+project-local file — see
+[`agent-user-schema.md § Global profile layer`](../../../docs/contracts/agent-user-schema.md#global-profile-layer-adr-138)
+for the merge rule. `init` and `update` operate on the **project-local**
+file only; `review` and `accept` cover BOTH layers' observation buffers
+(road-to-global-user-memory Phase 2/3); `show` renders the merged,
+effective profile by default, or the global layer's raw holdings with
+`--audit`; `delete` removes something the global learning channel wrote
+— an observation, a project's worth of observations, or a promoted
+profile field — with a tombstone in every case (Phase 4).
+
 **Why this is its own cluster:** `AGENTS.md` describes the *project*
 to the agent. `.agent-user.md` describes the *user* to the agent.
 Two distinct primitives — same `/agents` family for discoverability,
@@ -35,10 +47,11 @@ separate sub-commands for separation of concerns.
 | Sub-sub-command | Routes to | Purpose |
 |---|---|---|
 | `/agents user init` | `commands/agents/user/init.md` | Short interview → creates `.agent-user.md` |
-| `/agents user show` | `commands/agents/user/show.md` | Read-only render of the persona |
-| `/agents user review` | `commands/agents/user/review.md` | List buffered observations from `.agent-user.observations.jsonl` |
-| `/agents user accept` | `commands/agents/user/accept.md` | Apply a buffered observation with confirmation |
+| `/agents user show` | `commands/agents/user/show.md` | Read-only render of the persona; `--audit` renders the global layer's raw holdings |
+| `/agents user review` | `commands/agents/user/review.md` | List buffered observations from both the project-local and the global buffer |
+| `/agents user accept` | `commands/agents/user/accept.md` | Apply a buffered observation (either layer) with confirmation |
 | `/agents user update` | `commands/agents/user/update.md` | Open in IDE for manual edit; validate on save |
+| `/agents user delete` | `commands/agents/user/delete.md` | Delete an observation, purge a project's observations, or revoke a profile field — with a tombstone |
 
 Schema contract:
 [`docs/contracts/agent-user-schema.md`](../../../docs/contracts/agent-user-schema.md).

package/dist/agent-src/commands/check-current-md.md

@@ -109,16 +109,6 @@ After all fixes, re-run:
 ./scripts-run src/scripts/check_md_language <path> [<path> …]
 ```
 
-### 6. Mark hashes (only if `.agent-src.uncondensed/` was edited)
-
-For each modified source file:
-
-```bash
-./scripts-run src/scripts/condense --mark-done "{relative_path}"
-```
-
-This keeps `.condensation-hashes.json` consistent with the new content.
-
 ## Rules
 
 - **Never edit `.augment/` directly** — it's a generated projection.

package/dist/agent-src/commands/condense.md

@@ -23,8 +23,8 @@ packs:
 Condense agent config `.md` files from `src/` into token-efficient telegraph format
 and write the condensed output to `dist/agent-src/`.
 
-Uses SHA-256 hashes to track which source files changed since last condensation.
-Only changed files need recondensation — saving tokens and time.
+The projection is deterministic: `dist == rewrite(src)` byte-for-byte. A file is
+out of date only when its projection does not match that rewrite.
 
 ## Step 1: Sync non-.md files
 
@@ -41,99 +41,37 @@ changed `.md` files that need condensation.
 bash scripts/condense.sh --changed
 ```
 
-This lists only `.md` files whose source has changed since the last condensation (based on
-stored SHA-256 hashes). If no files changed → you're done.
+This lists only `.md` files whose projection is out of date (`dist != rewrite(src)`).
+If no files are listed → you're done.
 
 If you need to see ALL files regardless of change status:
 `bash scripts/condense.sh --list`.
 
-## Step 3: Condense each changed .md file
-
-**Core principle:** Condensation must improve execution quality, not just reduce length.
-A condensed file should be easier to scan, easier to trigger, and easier to execute than the original.
-
-For each changed `.md` file:
-
-1. Read the source from `src/{path}`
-2. Condense the prose using these rules:
-   - **Remove:** articles (a, an, the), filler (just, really, basically, actually, simply, essentially),
-     pleasantries, hedging, connective fluff (however, furthermore, additionally),
-     redundant wording, obvious framework knowledge, repeated explanations, decorative prose
-   - **Shorten:** "in order to" → "to", "make sure to" → "ensure", use short synonyms
-   - **Abbreviate** common terms when context is unambiguous: `DB`, `auth`,
-     `config`, `req`, `res`, `fn`, `impl`, `env`, `deps`, `ctx`. Skip if it
-     would be the first occurrence of the concept in the file, or if the
-     abbreviation collides with a domain term (e.g. `auth` stays
-     `authentication` inside an auth-module file). Never abbreviate inside
-     code blocks, frontmatter, file paths, command strings, or Iron Law
-     fenced blocks.
-   - **Arrows for causality:** `X causes Y` / `X leads to Y` / `X, then Y`
-     → `X → Y`. Keep arrows out of code blocks, frontmatter, and Iron Law
-     fenced blocks; only the surrounding prose uses them. (The example
-     phrases here are intentionally backticked so the inline-code
-     protection skips them — never strip those backticks.)
-   - **Fragments OK:** "Run tests before commit" not "You should always run tests before committing"
-   - **Drop:** "you should", "make sure to", "remember to" — state action directly
-   - **Merge** redundant bullets that say the same thing differently
-   - **Prefer** bullets over prose, direct instructions over explanatory paragraphs, one-line decisions over paragraphs
-3. **Copy-paste first, condense second:**
-   Before condensing ANY prose, extract and set aside these elements from the source.
-   They go into the condensed output **unchanged, byte-for-byte**:
-   - **All code blocks** (``` fenced or indented) — copy EVERY code block from source to output FIRST
-   - **YAML frontmatter** — copy verbatim
-   - **All inline code** (`backtick content`)
-   - **All URLs, links, file paths, commands**
-   - **All H1/H2 headings** — exact text preserved
-   - **Tables** (structure preserved, condense cell text only)
-   - **Technical terms, library names, API names**
-   - **Dates, version numbers, numeric values**
-   Then condense ONLY the prose around these preserved elements.
-4. **NEVER remove (even if verbose):**
-   - Trigger clarity (When to use / description)
-   - Decision hints that prevent mistakes
-   - Concrete validation checks
-   - Gotchas based on real failure patterns
-   - Anti-patterns that prevent recurring failures
-   - **Iron Law sections** — see "Iron Laws — do not touch" below
-5. **Enrich (SKILL.md files only):** During condensation, also improve agent-effectiveness:
-   - **Validation steps:** If a Procedure ends with a vague validation ("check if it works"),
-     replace with concrete checks (expected output, commands to verify, specific conditions)
-   - **Mini examples:** If no `Examples` section exists, add a short safe/unsafe or good/bad example
-     (2-4 lines max — not a tutorial)
-   - **Output format:** If the Output format section is vague ("explain everything"), sharpen it to
-     specific numbered expectations ("1. Code snippet 2. Where to register 3. Gotcha if relevant")
-   - **Anti-patterns:** If the same mistake keeps recurring, add a short anti-patterns section
-   - **Do NOT invent content.** Only concretize what the source implies. If the source says
-     "validate copy/paste safety", you can add "Check: no nested backticks, fully selectable".
-     But don't add unrelated sections.
-   - **Do NOT condense weak skills.** If the source has no procedure or no validation, fix structure first.
-   - **Reference skill:** See `src/skills/skill-writing/SKILL.md` for the gold standard
-6. Write the condensed output to `dist/agent-src/{path}`
-7. **MANDATORY: Run condensation quality check on this file:**
+## Step 3: nothing to condense — the projection is a deterministic copy
 
-```bash
-./scripts-run src/scripts/check_condensation --format text 2>&1 | grep "{path}"
-```
-
-If the output contains 🔴 (error) for this file: **STOP. Fix the condensed file before continuing.**
-Common errors and how to fix them:
-- `lost_code_blocks` → You dropped a code block. Copy ALL code blocks from source.
-- `modified_code_block` → Code block content changed. Replace with exact source content.
-- `frontmatter_mismatch` → YAML frontmatter differs. Copy verbatim from source.
+**ADR-201 (accepted 2026-07-29) removed the LLM prose rewrite from this command.**
+`.md` is copied verbatim and path-rewritten by `condense.sh --sync`; there is no
+per-file condensation step for the agent to perform any more.
 
-**Do NOT call `mark-done` until this file has zero 🔴 errors.**
+Why it was removed, measured with exact `tiktoken cl100k_base` over 429 artefacts:
+0 of them saved >= 500 tok, the aggregate saving was 0.86%, 267/429 pairs were
+byte-identical, and the 9 always-loaded kernel rules came out **36 tokens worse**.
+Determinism failed by construction — the hash covered the source and never the
+output, so divergence was undetectable, and it went undetected three times in a
+single session.
 
-8. Show word count: `{original} → {condensed} words ({saved}% saved)`
-9. **Mark as done:** `bash scripts/condense.sh --mark-done {path}`
+The instruction that used to live here already demanded *"copy EVERY code block
+from source to output FIRST, unchanged, byte-for-byte"*. It was clear, and it was
+broken anyway in six artefacts — corrupting template blocks users copy verbatim —
+because nothing checked it. That is the case for removing the rewrite rather than
+restating the rule more firmly.
 
-### Batch processing
+What survives is the one deterministic transform: `apply_path_rewriter` fixes
+relative links so they resolve from the delivered location (`../../docs/…` →
+`../docs/…`, ~38 artefacts). It runs automatically on every copy.
 
-When condensing multiple files, process them in batches of ~10.
-Mark each file done after writing it. After each batch, show a progress summary:
-
-```
-Batch 1/5 complete: 10 files, avg 42% saved
-```
+**If you were sent here to condense a file: don't.** Edit `src/`, run
+`--sync`, and let Step 4 verify that `dist == rewrite(src)` byte-for-byte.
 
 ## Step 4: Final verification gate
 
@@ -143,7 +81,7 @@ Run BOTH checks. Both must pass before finishing.
 bash scripts/condense.sh --check
 ```
 
-Must show ✅ (hashes in sync).
+Must show ✅ (`dist/agent-src/` in sync with source).
 
 ```bash
 ./scripts-run src/scripts/check_condensation
@@ -162,14 +100,6 @@ Read `verbosity.post_action_reports` from `.agent-settings.yml` (default
 - `full` → multi-line table with per-category stats (files condensed,
   avg savings).
 
-## Hash management
-
-- Hashes are stored in `.augment/.condensation-hashes.json` (committed to Git).
-- `bash scripts/condense.sh --sync` automatically cleans up hashes for deleted source files.
-- `bash scripts/condense.sh --mark-all-done` marks ALL current `.md` files as condensed
-  (useful after an initial full condensation or when bootstrapping the hash file).
-- A file with no stored hash is always treated as "changed".
-
 ## Iron Laws — do not touch
 
 Sections under headings matching `Iron Law`, `Iron Laws`, or `The Iron Law` (any
@@ -202,7 +132,9 @@ If an Iron Law section genuinely contains filler (rare): edit the SOURCE in
 
 ## Condensation quality checklist
 
-**Also apply the `preservation-guard` rule** — strongest validation, example, anti-pattern, and decision hints must survive condensation. Iron Laws are non-negotiable.
+**Also apply the [preservation-guard](../rules/preservation-guard.md) rule** — strongest validation, example, anti-pattern, and decision hints must survive condensation. Iron Laws are non-negotiable.
+
+See also: [markdown-safe-codeblocks](../rules/markdown-safe-codeblocks.md) for fenced-block hygiene.
 
 After condensing each file, verify:
 
@@ -237,4 +169,3 @@ Unsafe (DO NOT do this):
 - **Only write to `.augment/`** — the condensed output directory.
 - **Preserve ALL technical content** — only condense natural language prose.
 - **YAML frontmatter** in command/skill files must be preserved exactly.
-- **Always run `bash scripts/condense.sh --mark-done {path}`** after writing each condensed file.

package/dist/agent-src/commands/cost/report.md

@@ -37,6 +37,13 @@ Emits a Markdown summary of the session (model breakdown, per-tier cost,
 total USD) and appends a JSON record to
 `agents/cost-tracking/sessions.jsonl`.
 
+The summary now also reports a **replay dedup ratio** and splits **main session vs
+subagent legs**. Both matter for reading the number honestly: Claude Code replays
+the same assistant turn into many transcripts, so before the dedup fix this report
+overstated tokens and cost by roughly 2x — and the overstatement grew with
+conversation length. Any cost figure quoted from a run before that fix should be
+treated as pre-dedup.
+
 If `track.mjs` exits 2 with `no Claude Code project dir`, the user is not in
 a Claude Code session — surface that and stop.
 

package/dist/agent-src/commands/fix/portability.md

@@ -91,14 +91,6 @@ After all fixes, re-run:
 ./scripts-run src/scripts/check_portability
 ```
 
-### 6. Mark hashes
-
-For each modified file:
-
-```bash
-./scripts-run src/scripts/condense --mark-done "{relative_path}"
-```
-
 ## Rules
 
 - **Always fix in `src/`** — never edit `.augment/` directly.

package/dist/agent-src/commands/fix/refs.md

@@ -88,14 +88,6 @@ After all fixes:
 
 Show final result.
 
-### 6. Mark hashes
-
-For each modified file:
-
-```bash
-./scripts-run src/scripts/condense --mark-done "{relative_path}"
-```
-
 ## Rules
 
 - **Always fix in `src/`** — never edit `dist/agent-src/` or `.augment/` directly.

package/dist/agent-src/commands/humanize.md

@@ -35,7 +35,7 @@ counterpart to the write-engine's built-in step 4b audit.
 - `/humanize <path>` — read the file at `<path>` (read-only).
 - `/humanize --voice=<ghostwriter-slug|me>` — optionally resolve a voice
   via the write-engine's style-source resolution (§ 1 of the
-  [`write-engine`](../docs/contracts/write-engine.md) contract):
+  [`write-engine`](../../docs/contracts/write-engine.md) contract):
   a ghostwriter slug loads `agents/reference/ghostwriter/<slug>.md`, `me`
   loads `.agent-user.md`. The resolved fingerprint takes precedence over
   humanizer defaults (a voice that legitimately uses em dashes wins). No
@@ -96,7 +96,7 @@ never reworded, never dropped.
 
 - [`humanizer`](../../../src/skills/humanizer/SKILL.md) — the skill this
   command wraps.
-- [`write-engine § 4b`](../docs/contracts/write-engine.md) — the
+- [`write-engine § 4b`](../../docs/contracts/write-engine.md) — the
   built-in audit for ghostwriter/post-as drafts.
 - [`content-quoting-floor`](../../../src/rules/content-quoting-floor.md)
   — quoted text is never rewritten.