retro

Follow CLAUDE.md Writing Standards strictly. No hedge words, lead with insights/numbers, tables over prose, every sentence must earn its place.

Produce a retrospective that answers three questions in one artifact:

1. What was resolved? Addressed callouts with the user↔Claude exchange.
2. What is still open? Fresh callouts left unresolved in the window.
3. Did the market price the narrative? Per-ticker overlay of newsflow + earnings results/transcripts + price action. Trade ideas rank by the narrative-price delta — the gap between what official sources signaled and what the stock did. Biggest gaps are positioning opportunities; alignment is already-priced.

Output is an immutable Research note — one new note per run, never overwrite. Trade recommendations are informational (link-back + suggested follow-up skill); this skill never mutates conviction: or status: (Tier 3 Confirmation-Required per CLAUDE.md).

Execution context — subagent delegation (2026-07-08, MANDATORY)

Delegate the ENTIRE run (Step 0 pre-flight through Phase 9 output) to ONE Agent subagent (subagent_type: general-purpose, run_in_background: false). Pass this skill's full instructions plus the resolved window in the agent prompt. The subagent performs all reads, WebSearches, transcript fetches, lock acquire/release, and writes (the immutable Research note + Log back-references + _hot.md + _followups.md open-findings entries (Phase 7.4) + the calibration JSONL append/score (Phase 4.5)), and must END its final message with the complete Phase 9.2 user-facing report. The main thread renders that returned report verbatim in chat — never re-summarize it, never drop the Trade Ideas table.

Mental Models reading gate MUST cross the delegation boundary (MANDATORY — CLAUDE.md; _shared/mental-models-section.md). The subagent does NOT inherit CLAUDE.md, so the agent prompt MUST embed this gate verbatim for the Phase 4 trade-idea synthesis: "Before ranking trade ideas from the narrative-price gaps, read Mental Models/Generalist - Overview.md + the matching Mental Models/Industry - X.md/Lens - X.md for the tickers surfacing. Apply the READING PROTOCOL — the market-vs-vault gap is a hypothesis, not a verdict; run the base-rate adversarially (most gaps close because the market is right); an inverted-bear/inverted-bull signal is a disconfirm trigger on the vault's own stance, not automatic alpha."

Recursion guard — the agent prompt MUST also embed verbatim: "You are the EXECUTOR of this skill run, not a coordinator: do NOT re-delegate any part of this work to further Agent calls — the 'Execution context — subagent delegation' section of the instructions you were passed applies to the main thread only and is already satisfied by your existence. Perform all reads, searches, and writes yourself, and end your final message with the complete Phase 9.2 report." Without this line the passed instructions include the delegation mandate itself — an invitation to recurse.

This supersedes the 2026-06-07 "main-thread, no fork" decision (see Design notes) — that revert was specific to the frontmatter context: fork mechanism, which returned the report as unrendered stdout (blank panel). Agent-tool delegation is a different mechanism: the subagent's final message returns to the main thread as a tool result, and the main thread re-emits it in its own response, so rendering is preserved. This keeps the ~400K-token read/search budget out of main context (main receives only the final report). If a future run shows a truncated or missing report, the fallback is to remove this section and restore main-thread execution — the Research note persists to disk either way.

Arguments

$ARGUMENTS parses to a window keyword plus an optional trailing ticker (scoped mode, 2026-07-08).

Input	Canonical window	Day count
empty, 1w, w, week, weekly, 7d	1w	7
1m, m, month, monthly, 30d	1m	30
1q, q, quarter, quarterly, 90d	1q	90

Default window: 1w. A token that is not a window keyword is parsed as a ticker for scoped mode (/retro 1w NVDA, /retro NVDA → default 1w scoped to NVDA). Resolve the ticker to Theses/TICKER - *.md; if no thesis matches, reject: ❌ /retro no thesis for ticker "[input]" (and not a window keyword). Windows: 1w | 1m | 1q. A recognized window with an unrecognized SECOND token that is also not a thesis ticker → same reject.

Scoped mode (SCOPE_TICKER set) changes three things vs the vault-wide default:

1. Phase 1 reads only the scoped thesis + its _graph.md cluster peers' sector/macro links (not all ~60 files) — see Phase 1 scoped path.
2. TICKER_UNIVERSE = { SCOPE_TICKER } (∪ cluster peers if the graph primer is present and the user wants competitive context) — overrides the window-activity derivation in §2.4. The scoped ticker is analyzed even with zero window activity (the user asked for it explicitly), so the Phase 2.5 empty-window short-circuit does NOT fire in scoped mode.
3. Phase 3 issues the 3-channel query for the one ticker (fits a single message; the ≤5-ticker one-message path already covers it) — no ~126-query fan-out.

The vault-wide lock is retained even in scoped mode (Invariant 7 — retro still writes a Research note + may Log-append).

Step 0: Pre-flight (MANDATORY — runs before Phase 1)

0.1: Acquire vault lock

Acquire a vault-wide scope lock per .claude/skills/_shared/preflight.md Procedure 1. Timeout budget: 10 minutes (parallel reads across ~60 files + N parallel WebSearches).

Capture the token at Step 0.1, verify ownership (Procedure 1.5) at every subsequent Bash block, release in the final reporting Bash block via rm -f "$LOCK_FILE".

0.2: Rename-marker pre-flight (warn-only)

Glob .rename_incomplete.* at vault root. If any marker exists, emit:

⚠️ In-flight rename repair(s) detected: [list markers]. Retro scan will proceed but its Research note's wikilinks to the affected ticker(s) use current filenames. Complete rename repair before running downstream /sync.

DO NOT abort. Retro is a read-mostly exploratory operation; its output is a new Research note that does not mutate existing thesis bodies. Mid-rename state does not silently split propagation the way a vault-wide /sync would.

0.3: Window computation

# Example for 1m window
DAYS=30
TODAY=$(date +%Y-%m-%d)
WINDOW_START=$(date -v-${DAYS}d +%Y-%m-%d 2>/dev/null || date -d "${DAYS} days ago" +%Y-%m-%d)
echo "WINDOW|$WINDOW_START → $TODAY"

Record WINDOW_START, TODAY, and window label (1w|1m|1q) for Phase 6 frontmatter.

0.4: Graph primer (optional — read if present)

Read _graph.md per .claude/skills/_shared/graph-primer.md Mode A (ticker-scoped, applied in aggregate during Phase 4 per-ticker synthesis). If absent or unparseable, proceed without primer — log ℹ️ _graph.md absent — retro proceeds without cluster context. Never block.

The graph orients Phase 4 synthesis (which ticker is in which cluster, shared peers) — it never replaces thesis reads.

Phase 1: Activity-Gated Batch Read

Scoped mode (SCOPE_TICKER set): skip the triage — read only Theses/SCOPE_TICKER - *.md in full, plus (if the graph primer loaded) the sector/macro notes it wikilinks. Typically 1–4 files. Proceed to Phase 2.

Vault-wide mode — grep-triage first, then read only active files (2026-07-08): most theses have NO activity in a given window (especially 1w), so a blanket ~60-file full read is wasteful. Gate the reads:

1. One grep pass over Theses/*.md, Sectors/*.md, Macro & Technology/*.md for the three activity signals:
   • addressed callouts in window: > \[!(question|error|tip|todo)\].*→ Addressed 20\d\d-\d\d-\d\d
   • any fresh (unaddressed) callout: > \[!(question|error|tip|todo)\] 20\d\d-\d\d-\d\d without → Addressed on the line
   • Log date headers: ^### 20\d\d-\d\d-\d\d Emit filename + matched dates. (One grep -rnE with alternation, output_mode content — a single round-trip.)
2. Filter to the active set: a file is active if it has (a) an addressed-callout date within [WINDOW_START, TODAY], OR (b) any fresh callout (surfaced regardless of window — §2.2), OR (c) a Log date within the window. Fresh-callout and in-window-date matching is done in the reasoning layer from the grep output — no per-file read yet.
3. Read the active set in full, as ONE parallel tool-call batch. For a 1w window this is typically 5–15 files vs ~60; for 1q it grows toward the full set (correct — more activity to review). Do NOT serialize.

Files considered: Theses/*.md (every status — draft/active/monitoring/closed-but-not-archived all carry callouts + Log), Sectors/*.md, Macro & Technology/*.md. Research notes are NOT read here (they carry no callouts by convention; their Log is implicit via the owning thesis).

Why grep-gate, not section-target: callouts are inline anywhere in a body, so you cannot narrow within a file — but you CAN skip files with zero window activity entirely. The grep is the cheap discriminator; full-read is reserved for files that actually changed. TICKER_UNIVERSE (§2.4) is still derived from the active set's extracted callouts/Log, so gating out inactive files cannot drop a ticker that had activity.

Phase 2: Extraction (parse in-memory from Phase 1 reads)

For each file content, walk the body line-by-line. Track the current ## Heading at each line (parent-section detection walks back to the nearest ## — same approach as /archive-callouts Phase 2.2).

2.1: Addressed callouts in window

Match callout blocks with header:

> [!<type>] <fresh_date> → Addressed <addressed_date>[ <markers>]
> <body line 1>
...
> **Response:** <response line 1>
...

Qualify as in-window addressed iff WINDOW_START ≤ addressed_date ≤ TODAY. Extract per block:

- file: [path]
  ticker: [inferred from file — TICKER from "Theses/TICKER - Name.md", or null for sector/macro]
  parent_section: [nearest preceding ## heading]
  type: question | error | tip | todo
  raised_date: YYYY-MM-DD
  addressed_date: YYYY-MM-DD
  body: [joined text, `> ` prefix stripped]
  response: [joined text, `**Response:** ` prefix stripped]
  markers: [[[pinned]]] (informational only — not excluded; legacy [[preserve]] markers during deprecation transition also recorded)

Skip policy:

• Callouts inside ## Legacy Callouts — never parse (archive section).
• Callouts inside ## Log — never parse (defensive; Logs don't carry callouts by convention).
• Malformed headers (unparseable date, unrecognized type) — skip with ℹ️ Skipped malformed callout in [file]:[parent_section].

2.2: Fresh callouts in window

Match callouts with header > [!<type>] <fresh_date> (NO → Addressed token). Qualify as in-window fresh iff WINDOW_START ≤ fresh_date ≤ TODAY. Extract same schema as 2.1 with addressed_date: null and response: null.

Fresh [[pinned]] callouts ARE included — pinned means "do not address", not "do not report in retro". They still signal what the user was thinking about.

2.3: Log entries in window

Match headers ^### (YYYY-MM-DD)$ within any ## Log section. For each date header in [WINDOW_START, TODAY], collect every bullet line beneath it until the next ### or ## .

Extract per bullet:

- file: [path]
  ticker: [inferred from file]
  date: YYYY-MM-DD
  prefix: [first token-phrase before `:` if present — match against _shared/log-prefixes.md registry]
  text: [full bullet body]
  skill_origin: true | false  # matches any prefix in the skill-origin list

Skill-origin classification (matches _shared/log-prefixes.md skill-origin list): Stress test, Deepening, Deepened, ↳ CORRECTION: Deepened, Conviction reaffirmed, Status change:, CLOSED, Prune upgrade, Scenario , Initial thesis created, ROLLBACK to snapshot, Scenario REVERSED, Cross-thesis closure:, Cross-thesis closures:, Renamed file:, Comparison , Callout sweep:, Numbers refresh:, Cross-thesis signal via. (Transcript ingested: is deliberately NON-skill-origin per log-prefixes.md — transcript-driven activity counts as high-signal manual research.)

Sync-propagated entries are a third class, not manual signal. A bullet whose body STARTS with a [[Research/ wikilink under a date header annotated (/sync) is machine propagation output (sync Step 3f format: - [[Research/note]]: [delta] — [conviction impact]). Classify it sync-propagated: it counts toward TICKER_UNIVERSE (the ticker was genuinely active) but is weighted BELOW manual entries in VAULT_DIRECTION / §4.2 alpha-reflection analysis — otherwise a /sync all run reads as user conviction activity across every synced ticker.

Non-skill-origin, non-sync-propagated entries are manual or user-callout-driven activity — highest signal for retro ("what I actually thought about").

2.4: Ticker universe

Compute TICKER_UNIVERSE = unique set of tickers across 2.1 + 2.2 + 2.3 where ticker is non-null. This is the set of tickers that had ANY activity in the window and require market overlay.

Sector/macro-file callouts with no clear ticker attach are tracked separately in a portfolio_level_activity bucket — surfaced in the output but not ticker-keyed.

2.5: Empty-window short-circuit

If 2.1 + 2.2 + 2.3 = 0 entries across all files:

No activity found in [1w|1m|1q] window ([WINDOW_START] → [TODAY]).

Possible reasons:
  - No callouts addressed or raised in window
  - No /status, /sync, /stress-test, or manual Log edits in window
  - Window may be too short — try /retro 1m or /retro 1q

No Research note written. Lock released.

Release lock per Step 0.1 and exit cleanly.

Phase 3: Narrative + Price Overlay (parallel WebSearch + targeted fetches)

Core engine: per ticker, collect three signals — newsflow, earnings (if any landed in window), price action — then compute the narrative-price delta in Phase 4. Social sentiment is NOT collected (retail chatter, Twitter, Reddit all excluded per prior spec). The narrative channels here are OFFICIAL: press releases, SEC filings, earnings results and call transcripts, analyst upgrades/downgrades, major business news.

3.0: Local newsflow corpus (2026-07-20 — read before any WebSearch)

The n8n Workflow 3 sweep persists one story log per run at .data/news_stories/*.json ({date, stats, stories: [{title, cluster, score, sum, members}]} — sum is a factual per-story summary, members carry source URLs, ticker stories tag tk-<ticker> in member feedIds). If the folder exists, read the files whose date falls inside the retro window BEFORE issuing Phase 3 WebSearches:

1. Seed the News channel: per ticker, collect its window stories (match tk-<ticker> in members, or ticker/company name in title) with date + score + sum. These are dated, pre-scored, summary-carrying events — often richer than a WebSearch snippet. A ticker whose window newsflow is fully covered by corpus stories still gets its Price query (the corpus has no price data) but may skip or narrow its News query — note news_source: corpus in the extraction schema.
2. High-score / no-reaction signal: any corpus story with score ≥ 8 on a ticker whose window_price_move_pct lands within ±2% becomes a first-class Phase 4 input — material newsflow the market shrugged at is exactly the narrative-price delta this skill ranks. Tag these corpus-flagged so classification can cite the specific story (title, date, sum).
3. Absent or empty folder: skip silently — Phase 3 proceeds fully on WebSearch as before. The corpus is an accelerant, never a dependency.

3.1: Three-channel query plan per ticker

For each ticker ∈ TICKER_UNIVERSE, issue three distinct queries. Each channel surfaces a separate input to classification.

Channel	Query template	Extracts
Price	[TICKER] stock price [WINDOW_START] to [TODAY] percentage change	window_price_move_pct, intra-window volatility if visible
News	[TICKER] news [WINDOW_START] to [TODAY] analyst rating earnings guidance M&A	Major events (product launches, guidance updates, analyst actions, regulatory, M&A) — each with date + polarity (positive/negative/neutral)
Earnings	[TICKER] earnings [WINDOW_START] to [TODAY] results call transcript guidance	Results vs consensus, guidance direction, call-transcript key themes (only if an earnings event landed in the window)

3.1.5: Overlay cache check (2026-07-08 — skip re-querying fresh tickers)

Overlapping runs waste WebSearches: /retro 1w on Friday then /retro 1m the next Monday re-fetches identical recent price/news for every shared ticker. Cache the per-ticker overlay and reuse it within a freshness window.

mkdir -p .data/retro_cache      # gitignored (.data/)

For each ticker in TICKER_UNIVERSE, before issuing its 3 queries, check .data/retro_cache/${TICKER}.json. Reuse the cached overlay (skip all 3 queries for that ticker) iff BOTH:

• fetched_date is within 3 days of TODAY, AND
• the cached window_end ≥ this run's TODAY − 1 (the cached data already covers this run's recent edge).

Otherwise query fresh. After Phase 3 completes, write/overwrite .data/retro_cache/${TICKER}.json with {fetched_date: TODAY, window_end: TODAY, price_move_pct, news_events, earnings, dominant_polarity} (compact — the classification inputs only, not raw search text). The cache is advisory: a miss just means a normal query; staleness is bounded by the 3-day guard, so it never injects old market data into a fresh classification. Scoped mode uses the same cache for its single ticker.

3.2: Parallel batch issuance

Issue the three queries per non-cached ticker in the same parallel tool-call batch. For a 42-ticker universe with a cold cache this is ~126 WebSearches; chunk into messages of up to 25 WebSearch invocations per message, producing ~5-6 rounds. Never serialize within a chunk — only between chunks. Mirrors /catalyst Phase 2's pattern at 3× the fan-out. A warm cache (e.g. a 1m run days after a 1w run) drops this to only the tickers with new activity.

If the non-cached set is small (≤5 tickers) — including every scoped-mode run — all queries fit in one message.

3.3: Targeted fetches (earnings transcripts only)

Fetch selectively only when:

• An earnings event landed inside the window, AND
• The ticker already appears likely to rank in Top 5 by preliminary narrative-price gap (rough pre-screen on WebSearch snippets), AND
• The vault thesis has an active Log entry or addressed callout from the window (i.e., we care what the earnings said relative to our thinking)

Fetch order per qualifying ticker (2026-09-09 — Quartr first):

1. .data/transcripts/ cache — a TICKER_QN-YYYY[.quartr].json covering the window's earnings event costs zero fetches (a prior /transcript run already paid).
2. Quartr (per _shared/quartr.md §1–§3: ToolSearch-load, resolve companyId via the shared ID cache): mcp__quartr__get_event_summary on the window's earnings event is the cheap first read (structured summary, no paywall, ~1K tokens); escalate to mcp__quartr__read_transcript {section: "qna"} only when the summary leaves the narrative side of the gap ambiguous — Q&A is where the narrative lives, at roughly a third of a full transcript's size.
3. Legacy web path (Quartr unavailable or no coverage): if the Earnings WebSearch surfaced a canonical transcript URL (Seeking Alpha, company IR, Motley Fool), defuddle [URL] via Bash (strips navigation/comment noise) or WebFetch (fallback) — these pages are paywall/consent-wall-prone, which is why they are last.

Cap at 5 transcript-content pulls per retro run (summaries don't count toward the cap; full/qna reads and defuddle fetches do) — transcripts are long (~8-15K tokens each); unbounded fetches would blow the context budget. If >5 tickers qualify, fetch the top 5 by preliminary gap magnitude.

3.4: Extraction schema per ticker

ticker: TICKER

price:
  window_move_pct: [number | null if unavailable]
  direction: up | down | flat | unknown           # threshold: ±3% (see Phase 4)
  max_intraday_move: [if visible in snippets, else null]

news:
  events: 
    - date: YYYY-MM-DD
      headline: [text]
      polarity: positive | negative | neutral
      source: [publication or snippet origin]
  dominant_polarity: positive | negative | neutral | mixed | none
    # Aggregated across events: count polarities, take dominant; none = no material event

earnings:
  event_in_window: true | false
  date: YYYY-MM-DD | null
  surprise_direction: beat | miss | inline | null
  guidance_direction: raised | lowered | maintained | withdrawn | null
  transcript_fetched: true | false
  transcript_themes: [list of 3-5 key themes if transcript was fetched, else null]

source_snippets: [top 2-3 audit-trail snippets per channel]

If a channel returns nothing (thinly-traded ticker with no news, no earnings in window, etc.), record null / none — downstream Phase 4 handles null channels without aborting the ticker.

Phase 4: Three-Dimensional Classification

Per ticker, synthesize across three independent dimensions:

1. VAULT_DIRECTION — derived from Phase 2 extraction: strengthened | weakened | mixed | neutral

   • Aggregate addressed-callout resolutions + non-skill-origin Log entries. Classify each as strengthen/weaken/neutral. Dominant wins; ties → mixed; zero activity → neutral.

2. NEWS_DIRECTION — from Phase 3.4 news.dominant_polarity: positive | negative | neutral | none

   • If earnings event in window: roll earnings surprise_direction + guidance_direction into this bucket (beat + guidance raised → positive; miss + guidance lowered → negative; mixed → neutral).
   • none = no material news or earnings event surfaced.

3. PRICE_DIRECTION — from Phase 3.4 price.direction: up | down | flat | unknown

   • Threshold: up iff window_move_pct > +3%, down iff < -3%, else flat. Window-agnostic (1w, 1m, 1q all use ±3%) because absolute price moves — not annualized rates — determine whether the vault conclusion mattered by window end.

4.1: Primary signal — NARRATIVE-PRICE DELTA

The narrative-price delta is the core engine's output: compare NEWS_DIRECTION against PRICE_DIRECTION independent of vault stance.

News direction	Price direction	Delta label	Interpretation
positive	up	aligned-up	Narrative priced correctly; limited alpha
negative	down	aligned-down	Narrative priced correctly; limited alpha
positive	down	inverted-bear	Good news, stock fell — positioning/forward-risk signal
negative	up	inverted-bull	Bad news, stock rose — short-covering, capitulation floor, or forward-relief signal
positive	flat	unreactive-good	Priced in OR market skeptical of narrative
negative	flat	unreactive-bad	Priced in OR market dismissive of concern
none	up (>3%)	flow-bull	No news catalyst — positioning, sector rotation, or hidden signal
none	down (>3%)	flow-bear	No news catalyst — positioning, sector rotation, or hidden signal
none	flat	quiet	Nothing to say; ticker drops out of ranking
neutral	any	mixed	News was genuinely mixed; classification inconclusive
any	unknown	data-gap	Price unavailable; classification inconclusive

Gap magnitude — the ranking key in Phase 5:

gap_magnitude = 
  | inverted-bear  → 1.5 × |price_move_pct|   # highest-signal: good news sold into
  | inverted-bull  → 1.5 × |price_move_pct|   # highest-signal: bad news bought
  | flow-bull      → 1.0 × |price_move_pct|   # unsigned flow signal
  | flow-bear      → 1.0 × |price_move_pct|
  | mixed          → 0.5 × |price_move_pct| if |price_move_pct| ≥ 3% else 0   # a big move on genuinely-mixed news (e.g. beat + guidance cut) IS the market resolving the ambiguity — do NOT zero it out
  | unreactive-*   → 2.0 + 0.5 × |price_move_pct|   # base 2.0 + scaling: range 2.0–3.5; crosses the Phase 4.4 cluster threshold (3.0) at |move| ≥ 2%
  | aligned-*      → 0                         # already priced, no alpha
  | quiet/data-gap → 0

The 1.5× weight on inverted deltas reflects trader intuition: price rejecting narrative is a stronger signal than price confirming it. Confirmed narrative is consensus; rejected narrative is asymmetric information. The mixed 0.5× rung (added 2026-07-09) closes a gap where a large-move mixed-earnings name — the single most common ambiguous-catalyst case — scored 0 and silently dropped out of the Trade Ideas ranking despite being highly actionable. The unreactive-* formula was changed from max(2.0, 0.5×|move|) to 2.0 + 0.5×|move| (2026-07-09): with flat defined as |move| ≤ 3%, the old max() always evaluated to exactly 2.0 — dead code that made cluster-level unreactive signal (threshold 3.0) unreachable by construction.

4.2: Secondary signal — VAULT STANCE vs DELTA

Once the narrative-price delta is computed, the vault's recent stance determines whether we have alpha or reflection to do.

Delta	Vault direction	Read
inverted-bear	weakened	Vault was right, ahead of market — recent Log entries / addressed callouts weakened the thesis while market only now reacted. Harvest (trim, lock conviction).
inverted-bear	strengthened	Vault fighting market — vault strengthened view while market rejects positive narrative. Either vault sees through the surface catalyst (alpha) OR vault is wrong (reflection). Stress test.
inverted-bear	neutral / mixed	Missed signal — market moved on something vault didn't flag. Deepen the thesis; likely behind.
inverted-bull	strengthened	Vault was right, ahead of market — vault strengthened through bad news that market bought. Reinforce position.
inverted-bull	weakened	Vault fighting market — vault weakened while market bought the dip/news. Reassess; either vault too bearish or market wrong.
flow-bull / flow-bear	strengthened/weakened aligned	Vault anticipated flow — may be leading indicator. Consider action.
flow-bull / flow-bear	neutral	Vault uninformed — flow happening without vault view. Surface candidate for /thesis or /deepen.
unreactive-*	weakened/strengthened	Positioning/timing bet — vault has conviction, market deferring. Catalyst watch (see /catalyst).
aligned-*	any	No action — narrative and price agree; vault context is commentary, not alpha.

4.3: Output schema per ticker

ticker: TICKER
vault_direction: strengthened | weakened | mixed | neutral
news_direction: positive | negative | neutral | none
price_direction: up | down | flat | unknown
delta_label: aligned-up | aligned-down | inverted-bear | inverted-bull | unreactive-good | unreactive-bad | flow-bull | flow-bear | quiet | mixed | data-gap
gap_magnitude: [float]
read: [one-sentence interpretation from §4.2 matrix]
earnings_event_in_window: true | false
transcript_themes: [list or null]

4.4: Graph-primer aggregation

For each ticker with material classification (gap_magnitude > threshold of 3.0), consult the graph primer (if loaded in Step 0.4) for cluster peers. If ≥2 cluster peers share the same delta label → flag as cluster-level signal in the trade idea; otherwise idiosyncratic. Cluster-level inverted-bear across 3+ peers = sector rotation signal, not ticker-specific.

Phase 4.5: Calibration — persist this run's calls, score matured prior calls

Why (closes the self-agreement loop). Retro's core engine grades the market against the vault's own recent stance, and the vault's stance is derived by trusting prior Log/callout entries at face value (§4.2 VAULT_DIRECTION). Nothing checks whether the vault was actually right. Calibration is the independent check: it scores the retro's OWN past directional calls against realized price outcomes, building an empirical hit-rate over time — the vault's personal base-rate book ([[Generalist - Overview]] [G-10]: anchor to the reference class before the inside-view narrative). A retro that never scores itself is an inside-view engine with no feedback.

4.5a: Derive this run's directional calls

For each ticker in the Trade Ideas set (gap_magnitude > 0), derive an implied_forward sign from the §4.2 read + suggested action:

Suggested action (from §4.2)	implied_forward
Reinforce / add (inverted-bull + vault strengthened)	bullish
Harvest / trim (inverted-bear + vault weakened)	bearish
Deepen / New thesis / Monitor / Stress test	neutral (no directional bet — logged, but excluded from hit-rate; "stress test" = vault-fighting-market, the resolution is tracked separately in 4.5c)

flow-bull/flow-bear with an aligned vault direction → bullish/bearish respectively; unaligned → neutral.

4.5b: Persist (append-only JSONL)

mkdir -p .data          # gitignored

Append one line per Trade-Idea ticker to .data/retro-calls.jsonl (via printf >> — retro has no Write-to-arbitrary-path constraint here, printf is in allowed-tools):

{"call_date":"YYYY-MM-DD","window":"1w|1m|1q","ticker":"NVDA","delta_label":"inverted-bear","vault_direction":"weakened","implied_forward":"bearish","gap_magnitude":10.8,"scored":false}

Never rewrite existing lines here — this is the immutable call record. Scoring (4.5c) writes outcomes to a separate .data/retro-outcomes.jsonl keyed by call_date+ticker, so the call log stays append-only and a re-run cannot corrupt prior calls.

4.5c: Score matured prior calls

cat .data/retro-calls.jsonl (skip silently if absent — first ever run has nothing to score) and cat .data/retro-outcomes.jsonl (may be absent). In the reasoning layer, select calls that are:

• Mature: call_date is at least max(window_days, 21) days before TODAY (a directional call needs time to resolve), AND at most ~400 days old (bound lookback cost).
• Unscored: no matching call_date+ticker in retro-outcomes.jsonl.
• Directional: implied_forward ∈ {bullish, bearish} (neutral calls are not hit-rate-scored; "stress test" vault-fighting calls ARE resolved in 4.5c's separate tally).

Cap at 10 calibration price queries per run (bound WebSearch cost — same discipline as the Phase 3.3 transcript cap). Prioritise the oldest-unscored and highest-gap_magnitude calls. Reuse .data/retro_cache/ where fresh. For each selected call, one WebSearch: [TICKER] stock price [call_date] to [TODAY] percentage change. Classify realized direction (±3% threshold, same as Phase 4). Outcome:

• hit — realized direction matches implied_forward.
• miss — realized direction opposes implied_forward.
• push — realized |move| ≤ 3% (flat) or price data unavailable (data-gap, excluded from hit-rate denominator).

For "stress test" (vault-fighting-market) calls, resolve separately: did the price ultimately move the vault's way (vault vindicated) or the market's way (vault wrong)? Tally as vault_vindicated / market_right — this is the highest-signal calibration output, since vault-fighting-market is where the vault claims non-consensus edge.

Append outcomes to .data/retro-outcomes.jsonl:

{"call_date":"YYYY-MM-DD","ticker":"NVDA","implied_forward":"bearish","realized_move_pct":-9.4,"outcome":"hit","scored_on":"TODAY"}

Compute for the Phase 6 scorecard: overall hit-rate (hits / (hits+misses)), hit-rate by delta_label, and the vault-fighting-market vindication tally. If fewer than ~8 scored calls exist total, render the scorecard but caveat n too small — directional only.

Phase 5: Trade Ideas Ranking

Rank all tickers by gap_magnitude descending (from Phase 4.1). Top 3–5 populate the Trade Ideas section. Tickers with gap_magnitude = 0 (aligned, quiet, mixed, data-gap) drop out — they carry no trade signal.

Each trade idea renders:

• Ticker + one-liner: [delta_label] ([news summary] vs [price move]), vault [vault_direction]
• Read (from Phase 4.2 matrix): e.g., "Vault was right, ahead of market — harvest" or "Missed signal — deepen"
• Narrative detail: top 1-2 news events with polarity, earnings result if applicable, key transcript theme if fetched
• Price detail: window move %, intraday volatility if notable
• Suggested action (one of, mapped from §4.2 matrix):
  • Harvest / trim — vault was ahead, market just caught up
  • Reinforce / add — vault reading contradicted by surface, aligned with price direction
  • Stress test — vault fighting market, needs adversarial review → /stress-test TICKER
  • Deepen — vault missed the signal → /deepen TICKER [section]
  • Monitor for catalyst — vault conviction + unreactive price → /catalyst watch
  • New thesis candidate — flow move with no vault view → /thesis TICKER or /ingest
• Motivating evidence: wikilinks to the specific addressed callout(s) + Log entry(ies) + news headline that drove the classification
• Cluster flag: if Phase 4.4 surfaced cluster-level signal (≥2 peers same delta), prefix the read with [cluster signal] and list peers. Sector rotation trumps ticker-specific action.

Never auto-execute. Phase 7 optionally appends a Retro insight: Log entry to the affected thesis (see §7.1), but conviction and status changes require explicit user /status invocation.

Phase 6: Output — Research Note (immutable)

6.1: Filename

Research/YYYY-MM-DD - Retrospective [window-label] - Synthesis.md

Examples:

• Research/2026-04-24 - Retrospective 1w - Synthesis.md
• Research/2026-04-24 - Retrospective 1m - Synthesis.md

If a file with this exact name exists (same window already generated today), append numeric suffix: ... Synthesis 2.md, ... Synthesis 3.md. Each retro is a new immutable artifact (per user spec #3).

6.2: Frontmatter

---
date: YYYY-MM-DD
tags: [research, retrospective, synthesis]
status: active
source: vault retrospective
source_type: retrospective
window: 1w | 1m | 1q
window_start: YYYY-MM-DD
window_end: YYYY-MM-DD
tickers_touched: [TICKER1, TICKER2, ...]
propagated_to: []
---

Why propagated_to: []: retro notes are exploratory portfolio-level metadata, not per-thesis evidence. The body wikilinks reference many theses for context, NOT to claim each one needs a Log entry from /sync. The empty list is a terminal dedup signal to /sync Step 1 Check 2 — the producer skill (this /retro run) explicitly declares "no propagation needed." Without it, the next /sync would treat each body wikilink as a propagation target. Mirrors /surface's pattern.

Graph update deferred: _graph.md is owned exclusively by /graph. After /retro, run /graph last to register the retro research note in the dependency map.

6.3: Body structure

# Retrospective: [window-label] ([WINDOW_START] → [TODAY])

## Scope
- Window: [window-label] — [N] days
- Files scanned: [N] theses + [M] sectors + [K] macros
- Tickers with activity: [N]
- Addressed callouts: [N]
- Fresh (unresolved) callouts: [N]
- Log entries (non-skill-origin): [N]

## Trade Ideas (Top 5 by Narrative-Price Gap)
[ranked table — see Phase 5]

## Calibration Scorecard
[Phase 4.5c output — the vault's own track record on prior retro calls. Renders only if ≥1 prior call was scored this run; else "No matured prior calls to score yet."]

| Delta label | Scored | Hits | Misses | Hit-rate |
|---|---|---|---|---|
| inverted-bear | 6 | 4 | 2 | 67% |
| inverted-bull | 4 | 3 | 1 | 75% |
| flow-bull/bear | 5 | 2 | 3 | 40% |
| **Overall** | **15** | **9** | **6** | **60%** |

**Vault-fighting-market resolution** (the non-consensus-edge test): [N] "stress test" calls resolved — [X] vault vindicated / [Y] market right. [One-line read: is the vault's fighting-market instinct paying off, or is it fighting the tape?]

[Caveat if <8 total scored calls: "n too small — directional only."]

## Narrative vs Price vs Vault
[per-ticker table, sorted by gap_magnitude descending]

| Ticker | News | Earnings | Price | Vault | Delta | Gap | Read |
|---|---|---|---|---|---|---|---|
| NVDA | positive (guide raise, 2 upgrades) | beat + raised | -7.2% | weakened | inverted-bear | 10.8 | Vault ahead — harvest |
| BESI | positive (design win) | none | +0.4% | strengthened | unreactive-good | 2.0 | Positioning; catalyst watch |
| APP | none | none | +8.1% | neutral | flow-bull | 8.1 | Missed signal — investigate |
| AMD | negative (guide cut) | miss + lowered | +4.5% | strengthened | inverted-bull | 6.8 | Vault right — reinforce |

Column semantics:
- **News**: dominant polarity + 1-2 key events
- **Earnings**: surprise direction + guidance direction, or `none` if no event in window
- **Price**: window move %, threshold-flagged (`up`/`down`/`flat`) by ±3%
- **Vault**: direction derived from Phase 2 callout + Log aggregation
- **Delta**: label from Phase 4.1 narrative-price matrix
- **Gap**: `gap_magnitude` (ranking key)
- **Read**: one-phrase interpretation from Phase 4.2 vault-stance matrix

## What I Resolved
[per user spec #4 — proportional to window]

## What I Still Have Open
[fresh callouts with wikilink back to source file:section]

## Activity Log
[non-skill-origin Log entries, grouped by ticker then date descending]

## Portfolio-Level Activity
[sector/macro callouts + Log entries not keyed to a single ticker]

## Follow-up Skills
[concrete `/deepen`, `/stress-test`, `/status` commands the user can run next]

6.4: "What I Resolved" — proportional rendering (user spec #4)

Window	Rendering per addressed callout
1w	Full callout body + full response. Grouped by ticker, then by addressed date descending.
1m	Compressed: YYYY-MM-DD · [ticker] · [section] — [one-line summary of question] → [one-line summary of response key insight]. Full text archived via wikilink back to source file.
1q	Aggregated by ticker: TICKER: [N] callouts addressed across sections [list]. Top 2 insights: [one-line each, chosen by conviction-impact proxy — prefer entries that surfaced in Trade Ideas ranking]. Full text available via wikilink.

The rendering tier is a function of window length, not of callout count. A 1w retro with 40 callouts still renders full — the reader chose the tight window to get the full record.

6.5: "What I Still Have Open"

Fresh callouts in window, sorted by age descending (oldest fresh = most urgent):

- **[YYYY-MM-DD]** · [type] · [[file:section]] — [body one-liner]

Pinned fresh callouts render with · pinned marker — informational, signals user chose not to address.

6.6: Follow-up skills block

For each top-5 Trade Idea, emit the exact command(s) the user would type. Command selection is driven by the Phase 4.2 vault-stance-vs-delta matrix:

- [[NVDA - Nvidia]]: inverted-bear (news+, price -7.2%), vault weakened — harvest signal.
  - `/status NVDA conviction high→medium` — lock the read vault already made
  - `/brief NVDA` — generate IC-ready pitch for partial trim sizing

- [[APP - AppLovin]]: flow-bull (no news, price +8.1%), vault neutral — missed signal.
  - `/ingest https://[news-url]` — seed Research with what moved the tape
  - `/thesis APP` — vault has no view; decide whether to formalize

- [[AMD - Advanced Micro Devices]]: inverted-bull (guide cut, price +4.5%), vault strengthened — reinforce.
  - `/deepen AMD "Key Non-consensus Insights"` — articulate why vault sees through the miss

- [[BESI - BE Semiconductor]]: unreactive-good (design win, price +0.4%), vault strengthened — catalyst watch.
  - `/catalyst` — confirm BESI next catalyst date; position ahead

The retro never executes any of these — it lists them. The user decides which to run.

Phase 7: Thesis Log Entries (top 3–5 only)

7.1: Scope

Append one Retro insight: Log entry to the thesis of each ticker in the Trade Ideas Top 3 (NOT all top-5 — limit Log pollution to the highest-signal divergences). Tickers ranked #4-#5 are surfaced in the retro note only.

7.2: Format

### <TODAY>
- Retro insight: [1w|1m|1q] retro — [delta_label] (news [+|-|none], price [±X.X%]) vs vault [direction]. [One-sentence read from §4.2 matrix]. See [[Research/YYYY-MM-DD - Retrospective [window] - Synthesis]].

Example:

- Retro insight: 1w retro — inverted-bear (news +, price -7.2%) vs vault weakened. Vault was ahead of market on hyperscaler demand softening — harvest signal. See [[Research/2026-04-24 - Retrospective 1w - Synthesis]].

Prefix choice: Retro insight: is NOT in _shared/log-prefixes.md's skill-origin list (unregistered prefixes default to non-skill-origin). This is deliberate — per CLAUDE.md Workflow Rule 6, a non-skill-origin prefix forces the next /sync to treat the entry as research-driven and propagate normally to sector/macro notes. If the retro surfaces a divergence that materially changes sector-level thinking, /sync should pick it up.

7.3: Edit strategy (same-day append aware)

Use the same Case-4.3c.A / Case-4.3c.B two-path approach documented in archive-callouts/SKILL.md Phase 4.3c:

• If ### <TODAY> header already exists in the thesis's ## Log, append the bullet under the existing header. Never create a second same-day header; never reorder.
• If no ### <TODAY> header exists, append a new dated section at end-of-file (Log is always the last section in the thesis template).

After each Edit, re-grep for the new bullet to verify it landed. Silent-failure retry with expanded context, per Phase 4.3c retry rules.

If Log append fails after retry: do NOT abort the retro. Body Research note is the primary artifact; Log back-references are a convenience. Report ⚠️ Log back-reference to [[TICKER]] failed — retro note persisted, manual append recommended.

7.4: Open-findings register (per _shared/followups-contract.md)

The Retro insight: Log entries (7.1–7.3) live inside each thesis; _followups.md is the cross-run index that survives so an unactioned trade idea does not evaporate. For each Top-3 Trade Idea whose delta signals a vault-market mismatch requiring action (inverted-bear, inverted-bull, flow-bull/flow-bear with a vault view, unreactive-* with strong vault direction — i.e. every non-aligned top idea), append one entry to _followups.md ## Open:

- [ ] YYYY-MM-DD · retro · [[Theses/TICKER - Name]] · [delta_label] divergence, [suggested action] → user acts or dismisses · src [[Research/YYYY-MM-DD - Retrospective [window] - Synthesis]]

Dedup (followups-contract): grep ## Open for an existing retro · entry on the same thesis; if present, update its date rather than stacking (a weekly /retro must not pile the same idea). Read _followups.md first (create per schema if absent — retro has Write). Obey the 50-entry soft-cap warning. Resolvers /status (action taken) and /sync (absorbed into thesis body) close these entries later. Register-write failure is non-fatal, same as 7.3 — report and continue.

Phase 8: _hot.md Update

Read _hot.md then edit per _shared/hot-md-contract.md. If _hot.md does not exist, create it per CLAUDE.md Rule #9 schema.

8.1: Active Research Thread

• Same-topic continuation (existing thread already covers /retro work): append a dated line YYYY-MM-DD: [window] retro completed — top finding: [ticker + classification + key insight one-phrase].
• New topic: compress the outgoing thread into a single *Previous YYYY-MM-DD:* line, prepend, then replace thread body with the retro summary above.

8.2: Open Questions

For each Top-5 Trade Idea where the delta label signals a vault-market mismatch, add an Open Question. The label → question-template mapping:

Delta label + vault stance	Open Question template
inverted-bear + vault strengthened	[[TICKER]]: narrative+ but price fell ([X%]) over [window] while vault strengthened. What forward risk is the market pricing that vault missed? (flagged by /retro YYYY-MM-DD)
inverted-bull + vault weakened	[[TICKER]]: narrative- but price rose ([X%]) over [window] while vault weakened. Is the bear case already discounted, or is vault over-focused on the miss? (flagged by /retro YYYY-MM-DD)
flow-bull / flow-bear + vault neutral	[[TICKER]]: stock moved [X%] over [window] with no catalyst and no vault view. What is the market seeing that vault hasn't formed a view on? (flagged by /retro YYYY-MM-DD)
unreactive-* + vault strong direction	[[TICKER]]: vault [direction] on [narrative] but price flat over [window]. What catalyst is needed to close the gap? (flagged by /retro YYYY-MM-DD)

Dedup: suppress only if an existing question contains BOTH the ticker wikilink AND the distinctive fragment flagged by /retro. Existing questions from /thesis, /stress-test, /surface, /catalyst must NOT suppress this — retro-specific dedup is narrow by design. Same rationale as /catalyst Phase 5 dedup logic (§148 of catalyst SKILL.md).

aligned-* deltas do NOT generate Open Questions — market agreed with narrative, no puzzle to investigate.

8.3: Other sections

• Latest Sync: not touched (owned by /sync).
• Sync Archive: not touched (owned by /sync).
• Recent Conviction Changes: not touched (owned by /status).
• Portfolio Snapshot: not touched (owned by /sync).

8.4: Word cap

After edits, count total. Over 8,000 → run compression trigger order per hot-md contract §Compression trigger order (steps 1-3: drop Sync Archive / Previous: / merge OQ; steps 4-5: drop pre-14d OQ cohorts + roster-compress pre-30d RCC entries; step 6: aggressive RCC fallback if still over hard cap). Over 10,000 after step 7 → abort _hot.md update with: ❌ _hot.md exceeds hard cap (10,000 words) after full compression — retro note still persisted.

Retro note write is unconditional; _hot.md update is best-effort.

Phase 9: Release Lock and Report

9.1: Release

Final Bash block verifies ownership and releases:

LOCK_FILE=".vault-lock"
EXPECTED_TOKEN="<paste-token-captured-from-Step-0.1>"
if [ -f "$LOCK_FILE" ] && grep -q "token: $EXPECTED_TOKEN" "$LOCK_FILE"; then
  rm -f "$LOCK_FILE" && echo "=== LOCK RELEASED ==="
else
  echo "⚠️ Lock ownership check failed at release — skipping rm to avoid stealing another skill's lock."
fi

Runs unconditionally — success, empty-window, WebSearch failures, Log-append failures all reach this step.

9.2: Report to user

## /retro [window-label] complete

Window:           [WINDOW_START] → [TODAY] ([N] days)
Files scanned:    [N] theses / [M] sectors / [K] macros
Addressed:        [N] callouts resolved
Still open:       [N] fresh callouts (oldest: [date], ticker: [X])
Log activity:     [N] non-skill-origin entries across [N] tickers
Market data:      [N] tickers with price; [N] news events; [N] earnings events; [N] transcripts fetched
Calibration:      [N] prior calls scored — overall hit-rate [X]% ([hits]/[hits+misses]); vault-fighting-market [X vindicated / Y market-right]  (or "no matured calls yet")
Open findings:    [N] entries added to _followups.md (Top-3 divergences)

Top narrative-price gaps (see full Research note for details):
  1. [TICKER] — [delta_label] ([news polarity] vs [price move]) — [one-phrase read]
  2. [TICKER] — [delta_label] ([news polarity] vs [price move]) — [one-phrase read]
  3. [TICKER] — [delta_label] ([news polarity] vs [price move]) — [one-phrase read]

Cluster signals: [list if any surfaced in Phase 4.4, else "none"]

Research note: [[Research/YYYY-MM-DD - Retrospective [window] - Synthesis]]
Log back-references appended to: [list up to 3 tickers]
_hot.md: [summary of section updates, or "no updates" if nothing changed]

Next steps:
  → Run /graph last to register the retro note in the dependency graph
  → Review Trade Ideas and decide whether to act via /status, /stress-test, /deepen, /ingest, or /thesis

Runs in a delegated subagent (see Execution context at top) — the Phase 9.2 report is produced inside the subagent and re-emitted verbatim by the main thread, so the ~60 full-file reads + up to ~126 WebSearch results + up to 5 earnings-transcript fetches (~400K tokens) stay out of main-session context. The Research note persists to disk as the authoritative artifact regardless.

Invariants

1. Immutable output — each run creates a new Research note. Never overwrite an existing retro file; append counter suffix when filename collides.
2. No conviction/status mutations — Phase 5 Trade Ideas are informational. /status changes require explicit user invocation.
3. propagated_to: [] is terminal — retro notes are metadata, not evidence; do not trigger fanout /sync propagation.
4. Retro insight: is NOT a skill-origin prefix — deliberate choice: allows /sync to propagate downstream sector/macro updates if a retro-surfaced divergence materially shifts sector thinking.
5. Graph-primer is an orientation aid, never a filter — per _shared/graph-primer.md anti-patterns. Retro reads every thesis in full regardless of cluster membership.
6. Empty window exits cleanly — Phase 2.5 short-circuit releases the lock and exits without writing.
7. Lock is vault-wide — prevents concurrent /sync, /catalyst, /surface, /prune, /graph from racing on shared reads.
8. Calibration call log is append-only — .data/retro-calls.jsonl records each run's directional calls immutably; outcomes are written to a separate .data/retro-outcomes.jsonl keyed by call_date+ticker. Scoring never rewrites the call log, so a re-run cannot corrupt or double-count prior calls. Both are gitignored (.data/).

Failure modes and recovery

Failure	Recovery
Lock acquisition fails	Hard-block with standard /shared/preflight.md §1.4 message. User waits or force-unlocks after verifying stale.
File Read fails mid-Phase-1 batch	Continue with successful reads; report ⚠️ N files unreadable — retro based on partial vault state. Reduces signal fidelity, does not block output.
Price WebSearch fails for a ticker	Record price.direction: unknown; ticker classified as data-gap, drops to gap_magnitude = 0, excluded from Trade Ideas. Still appears in the Narrative vs Price table with price: data unavailable.
News WebSearch fails for a ticker	Record news.dominant_polarity: none; ticker uses price-only classification path (flow-bull / flow-bear / quiet). Reduced signal fidelity, still produces trade ideas if price move is material.
Earnings WebSearch fails for a ticker	Record earnings.event_in_window: false (defensive — we can't confirm). Ticker classified on news + price only.
Transcript fetch fails (defuddle / WebFetch)	Record transcript_fetched: false, transcript_themes: null. Earnings still contribute to news_direction via surprise + guidance fields. Note in the retro: ⚠️ transcript unavailable for [TICKER] — classification uses earnings result only.
All Phase 3 WebSearch calls rate-limited	Skill continues; Phase 4 classification emits data-gap for every ticker. Market overlay section shows ⚠️ Market data unavailable for this run — retro shows vault activity only, no market classification. Re-run /retro later. Vault-activity sections (Resolved / Still Open / Activity Log) still populate fully.
Phase 7 Log-append fails on a specific thesis	Report the specific file; retro note persists. Manual fix: append Retro insight: bullet per Phase 7.2 format.
_hot.md update fails hard-cap check	Retro note still persists. _hot.md unchanged. Report ⚠️ _hot.md at hard cap — skipped update. Manual cleanup needed.
Empty window (no activity)	Phase 2.5 short-circuit — no note written, lock released.
Lock ownership lost (LOCK_STOLEN)	Abort immediately. Report which Phase was in-flight and whether the retro note landed. If retro note partially written, leave it (reviewable); do NOT attempt cleanup.

Cadence (operational suggestion)

Pair with /cron (see the /schedule skill). Suggested defaults:

• /retro 1w — every Friday evening
• /retro 1m — first trading day of the month
• /retro 1q — first trading day of the quarter

Each run produces an independent immutable artifact. The /graph last reconciliation should follow each retro to register the new Research note.

Design notes

• Why vault-wide lock (not read-only): retro writes a Research note, appends Log entries to 3-5 theses, and updates _hot.md. These writes conflict with concurrent /sync or /catalyst. Read-only lock is insufficient.
• Scoped mode (added 2026-07-08): single-ticker retro (/retro 1w NVDA, /retro NVDA) is now supported — a trailing non-window token parses as a ticker (Arguments). It sets TICKER_UNIVERSE = {ticker}, reads only that thesis + its graph-linked sector/macro notes in Phase 1, issues one 3-query message in Phase 3, and does NOT empty-window-short-circuit (the user asked for this ticker explicitly). The vault-wide view remains the default for portfolio-level "recommend trades." The graph primer is orientation-only here too (Invariant 5) — scoped mode narrows reads, never analysis correctness.
• Activity-gated Phase 1 (added 2026-07-08): vault-wide runs grep-triage all theses/sectors/macros for in-window callout/Log activity first, then full-read only active files (~5-15 for 1w vs ~60). Callouts are inline-anywhere so per-file section-narrowing is unsafe, but skipping zero-activity files entirely is safe and scales the read cost with actual activity. TICKER_UNIVERSE still derives from the active set, so no active ticker is dropped.
• Overlay cache (added 2026-07-08): per-ticker price/news/earnings overlay cached to .data/retro_cache/ with a 3-day freshness guard, so an overlapping 1w→1m run doesn't re-fetch identical recent data. Advisory only — a miss is a normal query; the freshness guard prevents stale market data entering a fresh classification.
• Fork history — frontmatter fork reverted, Agent delegation adopted: 2026-06-07 the frontmatter context: fork mechanism was reverted because the harness returned the report as unrendered stdout, leaving the chat panel blank. 2026-07-08 the skill moved to Agent-tool delegation (see Execution context at top): the subagent does the ~400K-token read/search work and returns the full report as a tool result; the main thread re-emits it verbatim, so rendering is preserved AND main-context cost is just the returned report. The two mechanisms are distinct — do not "restore main-thread execution" on the assumption that delegation reintroduces the 2026-06-07 blank-panel bug; it does not. Revert path if delegation ever truncates: delete the Execution-context section, and the skill runs inline again (report renders, but ~400K tokens land in main context).
• Why narrative-price delta is the core engine: "what did the market say happened vs what did the stock do" is the classical trader signal for positioning opportunities. Aligned narrative + price = already priced, no alpha. Inverted narrative + price = market is pricing something the surface news isn't showing — positioning, forward risk, capitulation, or vault's non-consensus view is right. The vault's recent stance then separates alpha (vault predicted the divergence) from reflection (vault missed it). This is why the 1.5× weight on inverted deltas vs 1.0× on flow-only signals: rejected narrative is asymmetric information, confirmed narrative is consensus.
• Why exclude social sentiment: per user spec — official narrative channels (press releases, SEC filings, earnings, analyst actions) are structured, date-anchored, and polarity-classifiable. Social sentiment (X/Twitter/Reddit) is noisy, hard to polarity-classify at scale, and typically echoes official narrative with a lag. Including it would expand WebSearch fan-out 2-3× without proportional signal gain.
• Why cap transcript fetches at 5 per run: earnings transcripts are ~8-15K tokens each. An uncapped fetch on a heavy-earnings week (5+ tickers reporting) would blow the context budget. Pre-screen selects the 5 most likely to rank in Trade Ideas Top 5 — low-ranking earnings contribute via the Earnings WebSearch snippet (surprise + guidance direction) without the full transcript read.
• Why proportional rendering — per user spec. A 1q retro with 300 addressed callouts rendered in full would exceed the useful signal density of the note. Aggregation at long windows preserves the "what changed" signal without forcing the reader through 300 verbatim exchanges.