@windyroad/retrospective 0.26.0 → 0.26.1-preview.814

package/.claude-plugin/plugin.json

@@ -78,5 +78,5 @@
     }
   },
   "name": "wr-retrospective",
-  "version": "0.26.0"
+  "version": "0.26.1"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@windyroad/retrospective",
-  "version": "0.26.0",
+  "version": "0.26.1-preview.814",
   "description": "Session retrospectives that update briefings and create problem tickets",
   "bin": {
     "windyroad-retrospective": "./bin/install.mjs"

package/skills/analyze-context/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: wr-retrospective:analyze-context
-description: Deep context-usage analyzer. Runs richer heuristics than run-retro Step 2c — per-turn attribution, per-plugin decomposition, suggestion generation, policy-breach detection. Produces a markdown report at docs/retros/<date>-context-analysis.md with an HTML-comment trailer carrying the bucket-snapshot for delta-from-prior comparison. Auto-fires from run-retro Step 2c when the combined trigger holds (calendar-elapse >14 days OR delta >20% any bucket, once-per-day guard) per ADR-043 Amendment 2026-06-08; also user-invokable on demand.
+description: Deep context-usage analyzer. Runs richer heuristics than run-retro Step 2c — per-turn attribution, per-plugin decomposition, suggestion generation, policy-breach detection. Produces a markdown report at docs/retros/<date>-context-analysis.md with an HTML-comment trailer carrying the bucket-snapshot for delta-from-prior comparison. Auto-fires from run-retro Step 2c when the combined trigger holds (calendar-elapse >14 days OR a bucket changes >20% AND >10 KB, once-per-day guard) per ADR-043 Amendments 2026-06-08 / 2026-06-17; also user-invokable on demand.
 allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Skill
 ---
 
@@ -13,11 +13,11 @@ This skill is the **deep layer** of the two-layer design in **ADR-043** (Progres
 ## When to use
 
 - The user invokes `/wr-retrospective:analyze-context` directly.
-- **Auto-fired by run-retro Step 2c** when the combined trigger holds (calendar-elapse >14 days OR delta >20% any bucket since prior snapshot) and the once-per-day guard is not satisfied (no `docs/retros/<TODAY>-context-analysis.md` exists). Per ADR-043 Amendment 2026-06-08 (P295 settlement).
+- **Auto-fired by run-retro Step 2c** when the combined trigger holds (calendar-elapse >14 days OR a bucket's byte total changes by >20% AND by more than the 10 KB absolute minimum-delta floor since prior snapshot) and the once-per-day guard is not satisfied (no `docs/retros/<TODAY>-context-analysis.md` exists). Per ADR-043 Amendment 2026-06-08 (P295 settlement) + Amendment 2026-06-17 (P372 absolute-floor gate on the delta axis).
 - The user is preparing to trim context — e.g. before a release that introduces new hooks or skills, or after observing early compaction in long-running AFK loops.
 - The user wants a baseline snapshot at a known-good moment (e.g. immediately after a P091-cluster fix lands).
 
-**Auto-fires from run-retro Step 2c, silent in interactive and AFK modes.** Per ADR-043 Amendment 2026-06-08 + ADR-044 framework-resolution boundary, this skill auto-fires from the cheap layer when the combined trigger condition holds (calendar-elapse OR delta breach, capped at once per day). The skill never invokes `AskUserQuestion` — it writes a committed `docs/retros/<TODAY>-context-analysis.md` report and exits. AFK orchestrators read the resulting report on iteration close; the user reviews on return.
+**Auto-fires from run-retro Step 2c, silent in interactive and AFK modes.** Per ADR-043 Amendment 2026-06-08 + ADR-044 framework-resolution boundary, this skill auto-fires from the cheap layer when the combined trigger condition holds (calendar-elapse OR a delta breach clearing both the 20% and the 10 KB absolute floor, capped at once per day). The skill never invokes `AskUserQuestion` — it writes a committed `docs/retros/<TODAY>-context-analysis.md` report and exits. AFK orchestrators read the resulting report on iteration close; the user reviews on return.
 
 ## Output Formatting
 

package/skills/run-retro/SKILL.md

@@ -55,7 +55,7 @@ After reading the briefing tree, score every entry in `docs/briefing/*.md` to de
 |-------------|--------|
 | >= +3 | Promote to Critical Points candidate. The agent adds the entry to the Critical Points roll-up in `docs/briefing/README.md` during Step 3. |
 | 0 .. +2 | Keep in the topic file. No roll-up change. |
-| <= -3 | Route to the **delete queue**. These entries are surfaced for user confirmation in a single batched `AskUserQuestion` at the end of this step. |
+| <= -3 | **Silent removal** — no `AskUserQuestion`. The agent removes / trims the entry as part of Step 3 briefing curation and surfaces each removal in the Step 5 retro summary's Signal-vs-Noise Pass table with its score + ADR-026 citation; the user audits there and corrects via P078 authentic-correction if a removal was wrong (removals are reversible from git). See **Delete handling** below. |
 
 **Per-entry persistence format**: each briefing entry carries a trailing HTML comment block:
 
@@ -68,18 +68,9 @@ The comment block is appended to the list item (or heading) that contains the en
 
 **Classification ownership (policy-authorised per ADR-013 Rule 5)**: the agent owns silent classification. No `AskUserQuestion` is fired for individual entry promotions, demotions, or keep decisions. The agent applies the ADR-026 heuristic directly: entry cited in a tool call (or paraphrased in reasoning) during the session = signal; never loaded or loaded-but-unused = noise; ambiguous cases still classify but with a tentative flag the next retro resolves.
 
-**Delete queue confirmation**: after scoring all entries, if any entries have a score <= -3:
+**Delete handling — silent (no `AskUserQuestion`)**: entries scored <= -3 are handled SILENTLY. NO `AskUserQuestion` fires to confirm deletes, at any score band, in interactive OR AFK mode. This is the load-bearing direction every other authority already states: ADR-044's framework-resolution boundary lists "Briefing add / remove / rotate" as a framework-mediated NOT-an-`AskUserQuestion` surface; Step 3's "Removals are silent (P135 / ADR-044)" clause; CLAUDE.md MANDATORY P132 worked-example list (which names "run-retro Step 1.5 silent classification, Step 3 briefing removals"); and ADR-013 Rules 1/5. The agent classifies, the deletes are applied during Step 3 briefing curation, and each removal surfaces in the Step 5 retro summary's "Signal-vs-Noise Pass" table (each candidate with its score + the ADR-026 citation that led to the noise classification). The user audits the Step 5 table and corrects via P078 authentic-correction if a removal was wrong (removals are reversible from git). The AFK queue-and-continue surfacing is specified in the fallback paragraph below (P352).
 
-1. Present a single `AskUserQuestion` with `header: "Delete briefing entries?"` and `multiSelect: false`.
-2. The question body lists each delete candidate with its score and the ADR-026 citation that led to the noise classification.
-3. Options (up to 4 per prompt, sequential if > 4):
-   1. `Confirm all deletions` — description: "Remove all listed entries from their topic files."
-   2. `Delete selected only` — description: "The agent will present a follow-up with per-entry checkboxes."
-   3. `Keep all (defer to next retro)` — description: "Leave entries in place; scores remain unchanged."
-   4. `Review individually` — description: "Present each entry one at a time for keep/delete decision."
-4. If the queue is empty, skip the prompt entirely.
-
-If the user chooses `Delete selected only` or `Review individually`, present subsequent `AskUserQuestion` calls as needed, respecting the 4-option cap per ADR-013 Rule 1.
+Firing a batched `AskUserQuestion` to confirm deletes is the inverse-P078 lazy-deferral trap — sub-contracting a framework-resolved decision back to the user (lazy classification per the Step 2d Ask Hygiene Pass). It is superseded prose; do NOT reintroduce it.
 
 **Tier 1 budget guard**: if promoting all score >= +3 entries would breach the 2 KB / ~10-bullet Critical Points budget (ADR-040), promote only the highest-scored entries until the budget is met and surface the remainder as a budget-overflow advisory in the retro summary.
 
@@ -92,7 +83,7 @@ If the user chooses `Delete selected only` or `Review individually`, present sub
 - "Weaker evidence so defer to user pick" — the pass IS the evidence-gathering step. Citing "weaker evidence" without running the pass is circular.
 - Fabricating "wrap-mode vs mid-session" exemptions that do not exist in this SKILL — Step 1.5 fires uniformly regardless of when run-retro is invoked.
 
-Per ADR-044 framework-resolution boundary: Step 1.5 is mechanical (silent classification + AFK queue surfacing). The pass MUST emit a populated Signal-vs-Noise table in Step 5 (or, in interactive mode, the delete-queue prompt). An empty / "Deferred" emit without scan evidence is a Step 1.5 violation per P332. Same lost-observation hazard P148 captures: the silent skip has zero recovery affordance, while the populated table preserves the agent's judgement on disk for user audit.
+Per ADR-044 framework-resolution boundary: Step 1.5 is mechanical (silent classification + AFK queue surfacing). The pass MUST emit a populated Signal-vs-Noise table in Step 5. An empty / "Deferred" emit without scan evidence is a Step 1.5 violation per P332. Same lost-observation hazard P148 captures: the silent skip has zero recovery affordance, while the populated table preserves the agent's judgement on disk for user audit.
 
 **See also**: P148 (Step 4b Stage 1 anti-pattern driver — same class, different step); P332 (run-retro meta-surface recurrence driver).
 
@@ -248,12 +239,12 @@ Per **ADR-043** (Progressive context-usage measurement and reporting for retrosp
 4. **Auto-invoke the deep layer when the combined trigger holds (ADR-043 Amendment 2026-06-08, settles P295)**. After rendering the cheap-layer report, evaluate the combined whichever-comes-first trigger:
 
    - **Calendar-elapse**: the most recent `docs/retros/*-context-analysis.md` (lex-desc sort on date in filename) is older than 14 days, OR no prior report exists.
-   - **Delta-breach**: any bucket's byte total in the just-rendered report has changed by more than 20% versus the prior snapshot (HTML-comment trailer from step 2).
+   - **Delta-breach**: any bucket's byte total in the just-rendered report has BOTH (a) changed by more than 20% versus the prior snapshot (HTML-comment trailer from step 2) AND (b) changed by more than the absolute minimum-delta floor of 10 KB — i.e. `|current − prior| > 10240` bytes. Both conditions are required. The absolute floor suppresses negligible-delta noise: a small bucket (`project-claude-md`, `jtbd`) can trip the 20% rule on a sub-2 KB edit (e.g. `project-claude-md` 4277→5897 bytes = +37.9% but only +1620 absolute), which does not warrant the deep layer's committed-report + subagent cost. The 10 KB floor is `not estimated — chosen as initial value, reassess after 6 months of cross-project use` per ADR-026, symmetric with the 14-day / 20% grounding (ADR-043 Amendment 2026-06-17). The calendar-elapse axis still re-fires a large-but-stable bucket every 14 days regardless of delta, so the floor loses no coverage of the dominant context cost.
    - **Once-per-day guard**: if `docs/retros/<TODAY>-context-analysis.md` already exists (where `<TODAY>` is the current ISO date), treat the trigger as already-satisfied and skip auto-invocation. The snapshot artefact itself is the state; no new persistent state file is needed.
 
    When the trigger holds AND the once-per-day guard is not satisfied, invoke `/wr-retrospective:analyze-context` via the Skill tool. The deep layer is silent (never invokes `AskUserQuestion`) and produces a committed `docs/retros/<TODAY>-context-analysis.md` report. Record the auto-invocation outcome in the cheap-layer section: `Deep analysis auto-fired — see docs/retros/<TODAY>-context-analysis.md.` Identical behaviour in interactive and AFK modes per ADR-013 Rule 6 + ADR-044 framework-resolution boundary — auto-invocation is framework-resolved mechanical action, not a user-decided surface.
 
-   When the trigger does NOT hold (no calendar elapse AND no delta breach), emit a one-line note: `Cadence trigger inactive — next auto-fire when calendar-elapse >14 days OR delta >20% any bucket.` This is non-blocking and never a prompt.
+   When the trigger does NOT hold (no calendar elapse AND no qualifying delta breach — a delta breach qualifies only when it clears BOTH the 20% and the 10 KB floor), emit a one-line note: `Cadence trigger inactive — next auto-fire when calendar-elapse >14 days OR a bucket changes >20% AND >10 KB.` This is non-blocking and never a prompt.
 
    When the trigger holds but the once-per-day guard fires (today's report already exists), emit: `Cadence trigger holds; auto-fire skipped — docs/retros/<TODAY>-context-analysis.md already present.`
 

package/skills/run-retro/test/run-retro-signal-vs-noise.bats

@@ -74,13 +74,21 @@ setup() {
   [ "$status" -eq 0 ]
 }
 
-@test "run-retro: Step 1.5 delete queue uses single batched AskUserQuestion (ADR-013 Rule 1)" {
-  run grep -F 'Delete briefing entries?' "$SKILL_MD"
+@test "run-retro: Step 1.5 delete queue is silent — no AskUserQuestion (ADR-044 / P132; reconciled by P393)" {
+  # Step 1.5 deletes are SILENT (framework-mediated per ADR-044 "Briefing add/remove/rotate"
+  # + the CLAUDE.md P132 worked-example "run-retro Step 1.5 silent classification, Step 3
+  # briefing removals"). The superseded "single batched AskUserQuestion to confirm deletes"
+  # prose was a self-contradiction against Step 3's "Removals are silent" clause; reconciled
+  # to silent-delete by P393. This asserts the reconciled contract.
+  run grep -F 'Silent removal' "$SKILL_MD"
   [ "$status" -eq 0 ]
-  run grep -F 'Confirm all deletions' "$SKILL_MD"
+  run grep -F 'Delete handling — silent' "$SKILL_MD"
   [ "$status" -eq 0 ]
-  run grep -F 'Keep all (defer to next retro)' "$SKILL_MD"
+  run grep -F 'do NOT reintroduce it' "$SKILL_MD"
   [ "$status" -eq 0 ]
+  # negative: the superseded delete-confirmation prompt must NOT return
+  run grep -F 'Delete briefing entries?' "$SKILL_MD"
+  [ "$status" -ne 0 ]
 }
 
 @test "run-retro: Step 1.5 AFK fallback defers delete queue to retro summary (ADR-013 Rule 6)" {