菜单
Synchronize recent research learnings across all affected theses, sector notes, and macro documents. Use after a research session or when user says "sync", "propagate", or "update everything".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Sweep addressed callouts older than a threshold (default 180 days) into the target note's
Generate a concise 1-page investment brief from a thesis — the elevator pitch version. Use when user says "brief", "summarise thesis", "pitch me [TICKER]", "one-pager", or "investment memo".
Side-by-side competitive comparison of two or more companies. Use when user says "compare", "X vs Y", "which is better", or asks about competitive dynamics between specific companies.
Targeted deep research to fill a specific gap in an existing thesis. Use when user says "deepen", "flesh out", "expand on", "fill in", or specifies a thesis section to improve.
Rebuild the vault dependency graph (_graph.md) via the deterministic generator script. /graph forces a full rebuild; /graph last (run after every /sync) rebuilds only if the vault changed; /graph [N] is treated as last. The generator reads the vault and writes _graph.md directly, so the file never streams through a model response.
Ingest content into the vault as structured research notes. Accepts a URL, a local file path, or no arguments (batch-processes _Inbox/). Use when user shares a URL, says "clip this", "ingest", or "process inbox".
| name | sync |
| description | Synchronize recent research learnings across all affected theses, sector notes, and macro documents. Use after a research session or when user says "sync", "propagate", or "update everything". |
| model | opus |
| effort | max |
| allowed-tools | Read Grep Glob Edit Write Bash(date * find * cp * mkdir * touch * awk * sed * grep * git diff *) |
Propagate recent research insights across all affected vault documents. For each affected note, think through the full implications of new research on every substantive section — not just append links.
Design rationale in .claude/skills/sync/RATIONALE.md (§N.M anchors).
Shared contracts this skill consumes:
.claude/skills/_shared/preflight.md — vault locks, rename markers, name sanitization, section probes..claude/skills/_shared/wikilink-forms.md — the 5 canonical wikilink forms every matcher must cover..claude/skills/_shared/hot-md-contract.md — _hot.md compression budget and section ordering..claude/skills/_shared/log-prefixes.md — registry of Log-entry prefixes carrying cross-skill semantics..claude/skills/_shared/sector-resolution.md — thesis sector: → sector note resolution..claude/skills/_shared/mental-models-section.md — the ## Mental Models section schema, the MANDATORY /Mental Models reading gate, and the merge/idempotency write procedure./sync (default) and /sync all: acquire vault-wide scope lock per preflight Procedure 1. Timeout: 15 min (/sync all), 5 min (default)./sync TICKER: acquire ticker:TICKER scope lock. Timeout: 5 min.Capture the token at Step 0.1, verify ownership (Procedure 1.5) at every Bash block, release via rm -f "$LOCK_FILE" in the final reporting Bash block.
Step 0.2 (rename-marker glob), Step 1.1 (find -newer), and Step 1.2 (Read _graph.md) are independent read-only probes that fire AFTER lock acquisition and BEFORE any mutation. Issue them as a single parallel tool-call batch — one message containing the .rename_incomplete.* glob, the find -newer .last_sync Bash, and the _graph.md Read. Ownership verification on subsequent Bash blocks (per preflight §1.5) still runs on every mutation block; this batch only collapses three sequential read round-trips into one. On rare lock-collision edge cases where another skill is mid-release, the wasted work is negligible (milliseconds of find + one Read).
/sync TICKER: preflight Procedure 2. Hard-block if .rename_incomplete.TICKER exists./sync / /sync all: glob .rename_incomplete.*. Hard-block if ANY marker exists — these modes propagate across the vault; a single mid-rename ticker would split inbound wikilinks across old/new names./sync does NOT write to _graph.md. Graph maintenance is owned exclusively by /graph (run /graph last after this skill). /sync reads _graph.md as an input; never writes it.
Parse $ARGUMENTS:
/sync (default) — Graph-assisted targeted sync. Uses _graph.md if present; file-direct fallback only when graph returns empty for a given source (§2.2). Use after post-ingest propagation./sync all — Full brute-force sync via graph-cached two-pass triage. Use after major research sessions or monthly maintenance./sync TICKER — Ticker-scoped sync. Reads the thesis directly and reconstructs adjacency from outbound wikilinks. No graph dependency.Change detection uses .last_sync as the timestamp watermark. Touch policy:
| Mode | .last_sync behavior | .sync_all_fresh marker |
|---|---|---|
/sync (default) | Touch at end | — |
/sync all | Touch at end | Write at end (triggers /graph full rebuild) |
/sync TICKER | Do NOT touch (§6.1) | — |
First run: if .last_sync does not exist, create it with epoch timestamp: touch -t 197001010000 .last_sync. This applies to ALL modes (including ticker-scoped — epoch placeholder preserves baseline without advancing the watermark).
If $ARGUMENTS specifies files, use those. Otherwise:
find Research/ Theses/ "Macro & Technology/" Sectors/ -newer .last_sync -name '*.md'
If no changed files AND no files specified → report "Nothing changed since last sync" and stop (skip Step 2.9 manifest — no-op runs are pure read).
For each changed research note, the target set is the UNION of every thesis plausibly affected.
Primary resolution — graph-assisted: if _graph.md exists, look up the research note in each thesis's research: adjacency. Collect matches.
File-direct fallback (T7.2 — conditional): run ONLY when graph-assisted returned EMPTY for this research note OR _graph.md is missing/unparseable. Resolves via:
ticker: frontmatter matching Theses/[TICKER] - *.md[[Theses/...]] wikilinks in the research note bodytags: containing ticker-shaped tokens (all-uppercase, 1-5 chars) matching Theses/[TICKER] - *.mdWhen graph returned non-empty, trust the graph — see §2.2 for why. /lint #23 catches residual reverse-index drift.
Step 1.2 resolves research note → thesis targets. This step does the same for sector and macro targets — explicitly at Step 1 time, not lazily inside Step 4.-1 / Step 5.-1 gates. Symmetric with 1.2; prevents the cognitive shortcut where thesis-idempotent research notes cause Step 4 / Step 5 to be erroneously skipped via mental over-reach of the Step 2 source-dedup clause (see §1.11 for the 2026-04-22 failure case that motivated this step).
For each changed research note, compute two maps:
sector_targets_per_research_note[note]:
sector: frontmatter → resolve via .claude/skills/_shared/sector-resolution.md to a Sectors/*.md path._shared/wikilink-forms.md against [[Sectors/...]] / [[Sectors/name]] / [[Sectors/name.md]] — resolved via sector-resolution contract.Union (a) ∪ (b); dedup on resolved path.
macro_targets_per_research_note[note]:
[[Macro & Technology/...]] forms (5 canonical forms).source_type: scenario frontmatter OR tags: containing macro → body-scan for any Macro & Technology/ references (treat as macro-focused note even without explicit body wikilink).macro: frontmatter field if the note convention uses it.Union across (a)/(b)/(c); dedup on resolved path.
These maps are inputs to Step 4.-1 condition (ii) and Step 5.-1 condition (ii) as direct lookups — not re-derived at gate time. Empty set is a valid value (research note has no sector/macro relevance); the gate still fires its skip path correctly in that case.
Authoritative-guard boundary: Step 4.-1 / 5.-1 remain the authoritative skip gates. Step 1.2.5 only makes the gate's inputs explicit and first-class. Do not duplicate the gate logic here.
Reverse-index resolution (primary): Use _graph.md's Reverse Index: Macro → Theses table for each changed macro.
Body-grep fallback (T7.6 — one batched regex): ALWAYS runs to catch asymmetric adjacency (thesis wikilinks macro but macro doesn't reciprocate — §2.1). For a set of changed macro basenames {X, Y, Z}, issue ONE Grep:
Grep pattern='\[\[(Macro & Technology/)?(X|Y|Z)(\.md)?(#[^\]|]+)?(\|[^\]]+)?\]\]' path='Theses/' glob='*.md' output_mode='files_with_matches'
(Basenames regex-escaped; alternation built from the changed-macro set.) Output gives file → needs Grep match re-parse to identify which macro(s); a second pass on the returned file subset extracts specific basename matches. Two Grep calls total, regardless of changed-macro count. Dedup against reverse-index results.
Graph-assisted only. Sector notes are required by template to list ## Active Theses with every active thesis. /lint #2 catches missing MOC entries.
Before detailed idempotency work, filter out theses whose status is closed. ONE Grep vault-wide:
Grep pattern='^status: closed' path='Theses/' glob='*.md' output_mode='files_with_matches'
For every thesis in the closed set that also appears in a target set from 1.2/1.3/1.4:
⚠️ Skipped [TICKER] — status: closed but file remains in Theses/ (failed archive move). Complete: mv "Theses/[file]" "_Archive/[file]" → /graph last. Or reopen: /status TICKER status closed→active [rationale].
If a thesis listed in graph adjacency no longer exists on disk (archived since last graph rebuild):
⚠️ Skipped [filename] — file not found (likely archived). Run /graph last to refresh the dependency map.
Research notes with no resolvable target via 1.2:
ℹ️ [filename] — no propagation target found. Consider adding ticker: frontmatter or [[Theses/...]] wikilinks, or run /thesis [TICKER] to create the missing thesis.
For each changed research note, run ONE Grep pass to find every thesis that might already reference it — using the 5 wikilink forms from _shared/wikilink-forms.md. This single scan serves Case 2a, Case 2b, and Check 3 simultaneously (§1.5).
Per-research-note primary grep (base-name escaped for regex metacharacters):
Grep pattern='\[\[(Research/)?<base-escaped>(\.md)?(#[^\]|]+)?(\|[^\]]+)?\]\]' path='Theses/' glob='*.md' output_mode='files_with_matches'
Log-section confirmation (semantics preservation): the primary grep matches the wikilink ANYWHERE in the thesis file, but idempotency is defined by prior Log entries, NOT by wikilinks in ## Related Research (the user may have manually wikilinked without a Log entry, in which case /sync should still write the Log entry). Confirm in-Log presence via one Bash block across the primary match set:
# ESC_PATTERN is the same regex as the primary grep, escaped for awk.
for f in <files-returned-by-primary-grep>; do
awk '/^## Log/{in_log=1; next}
in_log && /^## /{exit}
in_log && ENVIRON["ESC_PATTERN_MATCH"] ~ $0{found=1}
END{if(found) print FILENAME}' "$f"
done
Alternatively (cleaner): use grep -n to get the line number of each primary-grep match, then intersect with the ## Log section's line range (extractable via grep -n '^## ' "$f" once per file). Either approach confirms "is match line within ## Log section" — files where the only match lies in ## Related Research or elsewhere are NOT in the confirmed-Log set.
The result — wikilink_match_set — is the set of thesis files with confirmed in-Log references to this research note. These are the already-propagated targets.
This two-pass structure keeps the token-saving win of T7.1 (primary grep is O(1) per research note vs old O(N) Read per candidate) while preserving the original Log-section-only idempotency semantics.
Now classify the research note's propagated_to: frontmatter:
Case 2a — Field absent: no producer-side dedup info. For each ticker in the Step 1.2 target set:
wikilink_match_set → treat as already-propagated; skip. Log: ℹ️ Log-history backfill skip: [TICKER] already has Log entry referencing [research-note]. propagated_to: absent; retry suppressed.propagated_to: frontmatter as a union of wikilink_match_set (resolved to tickers) + freshly-propagated tickers + any secondary-key-matched tickers. See Atomicity rule below.Case 2b — Field present and non-empty (e.g., propagated_to: [TICKER1, TICKER2]):
propagated_to: → skip. Log: ℹ️ Skipped [TICKER] — producer skill already propagated (in propagated_to).propagated_to:): for each, check wikilink_match_set. If present → skip (already propagated via another path). If absent → check Step 1.8 secondary keys; if any match → skip; else → propagate. Log: ➕ Augmented propagation: [TICKER] — not in propagated_to but resolved via [wikilink | tag | ticker frontmatter].Case 2c — Field present and empty (propagated_to: []): terminal skip for ALL targets. Log once: ℹ️ Skipped all propagation for [research-note] — empty propagated_to: signals producer-side terminal decision. Do NOT evaluate augmented targets. Producer must explicitly opt-out with [] to trigger this case (§1.6).
When a Step 1.2 target thesis is NOT in wikilink_match_set but the user may have renamed the research note (wikilink drift), check stable identity keys in order. Reads candidate thesis Log section plus at most one research-note frontmatter per candidate.
Key A — source: URL match: if the current research note's source: is a URL (not "vault synthesis"), scan the thesis's ## Log section for Log entries whose cited research notes share the same URL. Requires cross-read of candidate research notes' frontmatter. Match → treat as already-propagated.
Key B — date: + ticker: tuple match: for vault-synthesis notes without URLs (/scenario, /stress-test, /compare outputs). Scan thesis Log for entries citing Research/ notes with the same date: AND same ticker: tuple.
Key C — tags: intersection + date match: weakest. Current note's tags include ticker-shaped tokens overlapping thesis's ticker:, AND thesis Log has an entry citing a Research/ note whose tags: include the same token AND that entry's date matches current note's date:.
On any match: ℹ️ Skipped duplicate propagation for [TICKER] — Log references equivalent research note ([key type]: source URL | date+ticker | tags). Current: [[Research/current-name]]. Prior match: [detail].
Secondary keys fire rarely (only on rename/move events); their cost is bounded by the size of Step 1.2 target set minus wikilink_match_set.
propagated_to: updatesUpdate the research note's propagated_to: frontmatter ONLY after EVERY target thesis's Log entry has landed successfully for this research note in this run. If any Edit failed:
propagated_to: at all for this research note./sync retries the missing target via Case 2a flow.If every append succeeded, union existing propagated_to: with the newly-propagated tickers AND with any tickers found via Case 2a backfill scan. This moves Case 2a retries into Case 2b on future runs, eliminating the scan cost.
Atomicity scope: ONE research note. Partial success across multiple research notes in the same run is fine — each gets its own all-or-nothing decision.
Per-research-note append-failure reporting: list failed targets in Step 8 under propagated_to: update skipped.
Pass 1 — graph-cached triage (no thesis file reads):
_graph.md ONCE. Parse the Adjacency Index with status: and log_tail: fields..last_sync. Captures manual edits with no Log entry (§3.1). Always escalates; read full thesis in Pass 2.log_tail: shows recent conviction/status/reaffirm prefix within last 14 days.Pre-T7.3 fallback: if _graph.md lacks log_tail: fields, extract frontmatter + Summary + last 3 Log entries per thesis in one batched Bash call as a transient in-memory substitute. Full AWK implementation in RATIONALE §10 (transitional — only fires if a pre-T7.3 _graph.md is restored from backup).
Pass 2 — deep read for High-delta only (single parallel batch):
Issue ALL Pass 2 reads as a single parallel tool-call batch — one message containing every High-delta full-thesis Read, every Low-delta section-slice Bash, and every sector Read. Do NOT serialize per-thesis. A typical 15-25-file Pass 2 lands in one round-trip instead of 15-25 sequential rounds.
This typically reduces deep reads from ~58 files to 15-25 AND collapses those 15-25 into one round-trip. Combined with Pass 1 being a single batch, total Phase 1 cost is ~2 round-trips regardless of vault size.
Theses/TICKER - *.md prefix glob.
_Archive/: ⚠️ [TICKER] is closed — file lives in _Archive/. To reopen: /rollback TICKER → select pre-status snapshot → /sync TICKER.⚠️ No thesis found for [TICKER]. Run /thesis [TICKER] to create it.⚠️ Multiple Theses/ files match prefix "[TICKER]" — [list]. Resolve filename collision before re-running.status: closed but still in Theses/: ⚠️ [TICKER] has status: closed but file remains in Theses/ (failed archive move). ..._graph.md — file-direct is the only path in ticker mode):
[[wikilinks]] categorized by directory (Sectors/, Macro & Technology/, Theses/, Research/).Research/ for files with ticker: TICKER in frontmatter..last_sync watermark (timestamp-independent).Maximum reasoning effort here. For each source note read fully and analyse:
Source deduplication (ANALYSIS-SCOPE ONLY): if a thesis appears as a changed source AND research notes referencing that thesis are also in the changed set, treat the research notes as primary sources for Step 2 reading — skip the thesis's Log entries citing those research notes during analysis (they duplicate the research note's content).
This rule does NOT suppress Step 4 sector or Step 5 macro propagation. Sector and macro propagation are gated separately at §4.-1 and §5.-1 against the research note's own sector_targets_per_research_note / macro_targets_per_research_note sets resolved at Step 1.2.5. Source-dedup only changes HOW YOU READ sources during analysis; it does not change WHAT YOU ACT ON downstream.
Common mis-application (from 2026-04-22 failure, §1.11): concluding "all thesis targets are idempotent per Step 1.7 → nothing to propagate" and skipping Step 4/5 entirely. This is wrong. Thesis idempotency (Step 1.7) and sector/macro propagation (Step 4/5) are independent propagation paths. A research note that has already been fully propagated to its thesis targets may still have pending sector or macro propagation — always evaluate Step 4.-1 and Step 5.-1 against the research note's own target maps from Step 1.2.5, regardless of thesis-target idempotency.
Propagation quality depends on this step's depth.
Gates Step 4 / Step 5 — a thesis-only change driven by a skill that already handled downstream sector/macro state should not trigger redundant /sync re-edits.
For each changed thesis, read the most recent Log entry from _graph.md's log_tail: field (single-line lookup, zero file reads). If log_tail: is missing (pre-T7.3 graph) OR the thesis isn't in the graph, fall back to one bounded Bash call:
# Batch-extract most-recent Log bullet for up to N theses.
# CONVENTION: Log entries are append-only with newest at the BOTTOM
# (CLAUDE.md Tier 2 §2: "only append new dated entries at the end").
# Track the LAST date+bullet pair seen, emit in END.
# CRITICAL: do NOT `exit` on first bullet match — that returns the OLDEST
# entry and silently misclassifies theses whose initial Log entry is a
# skill-origin prefix (e.g., "Initial thesis created") but have since had
# research-driven entries appended (the AVGO/MRVL/CRWD/PANW failure mode
# observed 2026-04-26).
for f in Theses/TICKER1 - Name.md Theses/TICKER2 - Name.md; do
echo "=== $f ==="
awk '/^## Log/{in_log=1; next}
in_log && /^## /{exit}
in_log && /^### [0-9]{4}-/{date=$2; next}
in_log && date != "" && /^- /{last_date=date; last_bullet=$0}
END{if(last_date) print last_date" | "last_bullet}' "$f"
done
For each changed thesis:
mtime > .last_sync) AND most-recent Log entry matches a skill-origin prefix from _shared/log-prefixes.md (registry §5 Conviction reaffirmed, §6 Status change: conviction, §7 Status change:, §8 CLOSED, §9 Prune upgrade, §11 Initial thesis created, §12 ROLLBACK to snapshot, §13 Cross-thesis closure: / Cross-thesis closures:, §14 Scenario REVERSED, §15 Renamed file:, §16 Comparison , §17 Callout sweep:, §18 Numbers refresh:) AND no research note in the changed-file set resolves to this thesis.The skill-origin classification keys purely on the most-recent Log prefix. Its known failure (CLAUDE.md Workflow Rule #6; §1.11): the user manually edits a thesis body (Bull Case, Risks, Industry Context…) but forgets to append a Log entry, so the most-recent prefix is still a prior skill's — the edit is misclassified skill-origin and sector/macro propagation is silently skipped. Previously this surfaced only as a passive Step 8 report line; now it is an active gate.
For each thesis about to be classified skill-origin, detect whether its body changed outside ## Log since the last sync:
git diff --unified=0 HEAD -- "Theses/TICKER - Name.md" (the vault is a git repo). Determine whether any changed hunk falls outside the ## Log section — parse hunk line ranges against the ## Log start line (awk: line number of ^## Log). Changes only within ## Log → genuinely skill-origin (skill Log-append). Any change above ## Log (or to a non-Log section) → body edit.git unavailable): diff the current body-outside-## Log against the most-recent prior (pre-sync …) snapshot for this thesis in _Archive/Snapshots/ (awk-extract everything before ^## Log from both, compare). No prior snapshot → cannot determine; retain skill-origin classification and rely on the Step 8 report (current behavior — no regression).On body-edit detection, do NOT silently reclassify and do NOT silently skip. Surface an inline prompt:
⚠️ [TICKER] classified skill-origin (most-recent Log prefix "[prefix]"), but its body changed
outside ## Log since the last sync — likely a manual edit with no Log entry (Workflow Rule #6).
Propagate this thesis's changes to sector/macro notes as research-driven? [Y/n]
(Y = treat as research-driven and propagate — recommended; n = keep skill-origin, skip sector/macro)
Default Y (mirrors the Mixed → research-driven precedent). On Y, move TICKER out of skill_origin_theses and process it through Steps 3/4/5 normally; also recommend the user append a Manual edit: Log entry so future syncs classify it correctly without the prompt. On n, leave it skill-origin. Batch the prompt if multiple theses trip it (one consolidated question listing all affected tickers).
Cost containment: the git diff runs ONLY for theses that would otherwise be classified skill-origin (typically 0–few per sync), never for research-driven theses. False-positive guard: if the user hasn't committed since a prior in-session skill edit, git diff HEAD may include that skill's own body edits — the prompt (not silent reclassification) is what makes this safe, since the user confirms.
Maintain in-memory accumulator skill_origin_theses: [TICKER, ...]. Steps 4 and 5 consult this before per-sector / per-macro propagation. The Body-change override runs before this accumulator is finalized, so any prompted-Y thesis is already removed.
The skill-origin enumeration above is a view of _shared/log-prefixes.md's skill-origin entries. /lint #29 verifies registry/enumeration alignment. When a new skill-origin prefix is added: update the registry first, then this enumeration.
Mandatory BEFORE any vault state mutation. Runs once per non-no-op run, immediately before Step 3.
Generate HHMMSS via date +%H%M%S once at the start of Step 2.9. This is THE canonical batch ID — Step 3c reuses it for all Tier A snapshots; Step 7.5 flips the same manifest.
Batch ID format:
/sync (default) and /sync all: sync-YYYY-MM-DD-HHMMSS/sync TICKER: sync-TICKER-YYYY-MM-DD-HHMMSS — prevents collisions between simultaneous /sync NVDA and /sync AMATSkeleton path:
_Archive/Snapshots/_sync-manifest (sync-YYYY-MM-DD-HHMMSS).md # default / all
_Archive/Snapshots/_sync-manifest (sync-TICKER-YYYY-MM-DD-HHMMSS).md # ticker-scoped
---
type: sync-manifest
batch: sync-YYYY-MM-DD-HHMMSS
mode: default | all | ticker-scoped
status: in-progress
date: YYYY-MM-DD
---
# Sync Batch Manifest (in-progress)
Manifest written at Step 2.9 before any file modifications. Populated at
phase boundaries (end of Step 3, 4, 5) then flipped to status: completed
at Step 7.5.
If this file persists with status: in-progress, the sync crashed or was
interrupted. Recovery: inspect the sections below, then /rollback to
restore any Tier A snapshots listed.
## Theses with snapshots taken (Tier A)
_populated at end of Step 3_
## Theses with Log-only appends (Tier B)
_populated at end of Step 3_
## Sector notes touched
_populated at end of Step 4_
## Macro notes touched
_populated at end of Step 5_
## Source research notes processed
_populated at end of Step 3_
❌ Sync manifest skeleton write failed: [path]
Reason: [error]
No vault state has been modified yet. Fix the filesystem issue (disk
space, permissions, concurrent writer) and re-run /sync.
Exit. Do NOT proceed.
Manifest writes happen at 4 discrete points — NOT per event. This is the T7.4 optimization: ~4 Edits on the manifest instead of ~49 for a typical /sync all run. In-memory accumulators per Step 7.5d collect events during each phase; phase-end Edit flushes all accumulated entries at once.
| Checkpoint | Fires at | Writes |
|---|---|---|
| Checkpoint A | End of Step 3 (all theses done) | Populates "Theses with snapshots", "Theses with Log-only appends", "Source research notes processed" |
| Checkpoint B | End of Step 4 (all sectors done) | Populates "Sector notes touched" |
| Checkpoint C | End of Step 5 (all macros done) | Populates "Macro notes touched" |
| Step 7.5 flip | After all writes complete | Flips status: in-progress → status: completed + completed_date: |
See §5.3 for crash-recovery implications. The phase-boundary design accepts intra-phase granularity loss (mid-Step-3 crash means that phase's theses appear only via snapshot file glob, not via manifest). /rollback remains correct in both cases.
Read the thesis with scope driven by Step 3b's impact map. Procedure:
# Example: read frontmatter + Summary + Bull Case + Risks for one thesis
awk '
/^---$/ && !fm_done { in_fm=!in_fm; print; if(!in_fm) fm_done=1; next }
in_fm { print; next }
/^## Summary/,/^## /{if($0!~/^## /||$0~/^## Summary/) print}
/^## Bull Case/,/^## /{if($0!~/^## /||$0~/^## Bull Case/) print}
/^## Risks/,/^## /{if($0!~/^## /||$0~/^## Risks/) print}
/^## Log/{in_log=1}
in_log{print}
' "Theses/TICKER - Name.md"
Always include frontmatter (for conviction/status) and full ## Log section (for drift detection in Step 3e). The Log is the expensive part — its tail is what drift analysis needs.
When section-scoping is skipped (read full):
/sync TICKER mode — always read in full (ticker-scoped cost is already bounded)./sync all (body may have arbitrary user changes).Map Step 2's insights against each thesis section. Be selective.
| Section | Update when... |
|---|---|
| Summary | Core investment case shifted — change in thesis direction or key assumption |
| Non-consensus Insights | New evidence strengthens/weakens existing insight, OR genuinely new angle |
| Outstanding Questions | Question answered (mark resolved + date), OR new questions raised |
| Business Model | Structural change to revenue model, product mix, go-to-market, segment economics |
| Industry Context | Competitive positioning, market share, pricing power shifted |
| Bull Case | New supporting evidence, or bull scenario element validated/invalidated |
| Bear Case | New counter-evidence, or bear scenario element validated/invalidated |
| Catalysts | New catalyst identified, or existing catalyst resolved — mark outcome |
| Risks | New risk emerged, existing risk probability changed materially, or risk retired |
| Conviction Triggers | New evidence validates/invalidates a trigger, OR threshold needs updating |
| Mental Models | A new research-activated /Mental Models trigger fires, an existing trigger's read materially changes, or models newly agree (disconfirm check) — per _shared/mental-models-section.md. High selectivity: most syncs leave it untouched. |
Do NOT update sections where the delta is trivial. Thesis contains synthesised conclusions, not duplicated research.
Mental Models section — before writing it, read the relevant /Mental Models files and apply the READING PROTOCOL per _shared/mental-models-section.md (MANDATORY reading gate). Merge new triggers as hypotheses-to-test; never overwrite; idempotent on model + trigger name. An unchanged section is the correct outcome for most runs.
Classify edits by tier:
If any Tier A edits planned, snapshot ALL Tier A theses in one Bash block using background shell jobs + wait (mirrors /compare 5.5a and /rollback Step 4 patterns — each cp is independent; sequential execution over N theses wastes N-1 round-trips):
mkdir -p _Archive/Snapshots
# For each thesis T in tier_a_thesis_snapshots (from 3g accumulator), fire cp in background:
cp "Theses/TICKER1 - Name1.md" "_Archive/Snapshots/TICKER1 - Name1 (pre-sync YYYY-MM-DD-HHMMSS).md" &
cp "Theses/TICKER2 - Name2.md" "_Archive/Snapshots/TICKER2 - Name2 (pre-sync YYYY-MM-DD-HHMMSS).md" &
cp "Theses/TICKER3 - Name3.md" "_Archive/Snapshots/TICKER3 - Name3 (pre-sync YYYY-MM-DD-HHMMSS).md" &
wait
# Check exit status of background jobs before proceeding
For single-thesis Tier A (/sync TICKER or default sync with one affected thesis), this collapses to a single foreground cp. The batch form only matters for multi-thesis /sync all runs.
Then issue all frontmatter-addition Edits as a single parallel tool-call batch (one Edit per snapshot, all in the same message):
snapshot_of: "[[Theses/TICKER - Company Name]]"
snapshot_date: YYYY-MM-DD
snapshot_trigger: sync
snapshot_batch: sync-YYYY-MM-DD-HHMMSS
Batch ID reused from Step 2.9. Proceed with edits to the ORIGINAL.
~~Original question~~ → Resolved YYYY-MM-DD: [answer].Key Metrics: update table if new numbers.
Related Research: add wikilink to new research note.
Conviction trigger check: if new data satisfies a predefined trigger: ⚡ Trigger hit: [quote the trigger] — [what happened].
Conviction drift detection — see _shared/log-prefixes.md for authoritative prefix list. /lint #29 verifies consistency. Scan Log backward from most recent, excluding entries that begin with (unconditional exclusions):
"Stress test" (registry §1)"Deepening" (registry §2)"Cross-thesis closure:" / "Cross-thesis closures:" (registry §13)"Scenario REVERSED" (registry §14)"Renamed file:" (registry §15)"Callout sweep:" (registry §17)"Numbers refresh:" (registry §18 — a metrics refresh is hygiene, not conviction sentiment; without exclusion each refresh consumes a drift-window slot)Conditionally exclude entries beginning with "Deepened" or "↳ CORRECTION: Deepened" within 14 calendar days of a "Stress test" entry (registry §3-§4; deepened_exclusion_days in .drift-config.md, default 14).
Anchor handling: if "Conviction reaffirmed" (registry §5) or "Status change: conviction" (registry §6) found, anchor the window there — only count entries after. If fewer than 3 entries after anchor, drift detection has insufficient data. If no anchor, use last 5 non-excluded entries.
Threshold (tunable via .drift-config.md — schema in INFRASTRUCTURE.md §2.7):
Base: 3/5 weakening (if conviction: high) or 3/5 strengthening (if conviction: low).
Post-stress-test suppression: if "Stress test" entry within last 30 days, raise to 4/5.
Missing config → defaults. Malformed → ⚠️ .drift-config.md malformed — using defaults.
conviction: high + threshold weakening → ⚠️ Conviction drift — [N]/[window_size] recent updates flagged headwinds ([threshold] required, post-stress-test suppression: yes|no). Reassess. To acknowledge: /status TICKER reaffirm [rationale]
conviction: low + threshold strengthening → 📈 Positive drift — [N]/[window_size] recent updates supportive. Reassess. To acknowledge: /status TICKER reaffirm [rationale]
Append to Log:
### YYYY-MM-DD
- [[Research/source-note]]: [what changed] — [conviction impact: unchanged/strengthened/weakened + 1 reason]
Max 2 lines. Log is an audit trail; analysis lives in the updated sections above.
For each thesis processed in Step 3:
(path, snapshot_path) to tier_a_thesis_snapshots.(path, log_entry_text, reason) to tier_b_thesis_log_appends where reason is cross-thesis propagation | augmented target | other.source_research_notes (dedup at checkpoint time).Checkpoint A fires once at end of Step 3 — one Edit populating all three manifest sections from the accumulators.
Consult skill_origin_theses from Step 2.5. A sector's "affected thesis set" is every changed thesis whose sector: frontmatter resolves to this sector (via _graph.md reverse index OR _shared/sector-resolution.md fallback).
Gate rule: skip Step 4 entirely for this sector ONLY IF BOTH conditions hold:
skill_origin_theses, ANDsector_targets_per_research_note (computed at Step 1.2.5).If either condition fails, proceed to Step 4.0. Log (skip path): ℹ️ Skipped sector update [Sector Name] — all [N] affected thesis changes are skill-origin AND no co-changed research note resolves to this sector per Step 1.2.5.
Condition (ii) is NOT satisfied by thesis-target idempotency (Step 1.7 Case 2a/2b/2c). Idempotent thesis propagation and empty sector target set are independent facts. The 2026-04-22 failure case collapsed them incorrectly — see §1.11.
Mixed set: at least one research-driven thesis OR research note resolves to sector → proceed per §4.2.
Before snapshotting, check if this sector already received propagation for the source research note today.
## Related Research section. Check for the source research note's wikilink (any of 5 forms per _shared/wikilink-forms.md).
## Log section. Scan for ### YYYY-MM-DD entry where entry_date == today AND entry references the source research note's wikilink.
ℹ️ Skipped sector update [Sector Name] — Log already contains today's entry for [research-note] (likely propagated by earlier /sync TICKER in this session).If edits modify existing analytical text (competitive dynamics, value chain, sector observations, comparison narratives), snapshot ALL Tier A sectors in one Bash block using background shell jobs + wait:
# For each sector in tier_a_sector_snapshots, fire cp in background:
cp "Sectors/Sector A.md" "_Archive/Snapshots/Sector A (pre-sync YYYY-MM-DD-HHMMSS).md" &
cp "Sectors/Sector B.md" "_Archive/Snapshots/Sector B (pre-sync YYYY-MM-DD-HHMMSS).md" &
wait
Then issue all frontmatter-addition Edits as a single parallel tool-call batch.
Add to each snapshot's frontmatter:
snapshot_of: "[[Sectors/Sector Name]]"
snapshot_date: YYYY-MM-DD
snapshot_trigger: sync
snapshot_batch: sync-YYYY-MM-DD-HHMMSS
Skip snapshot if only adding wikilinks.
/sync consideration and — if the sector's priced-in read shifted — update the sector's ## Investor heuristics (what consensus believes / where it's wrong: the vault's stated edge). Do NOT silently leave peers stale because the research arrived under the source ticker.## Mental Models section when new cross-company evidence activates or retires a sector-level model trigger, or changes its read — read the relevant /Mental Models files first and merge per _shared/mental-models-section.md (high selectivity; most syncs leave it untouched).Inspect each Edit tool call's return value (the post-edit content snippet it emits). The Edit tool reports success iff the replacement landed, and the return value shows the new content; a separate re-read adds a round-trip without adding information. Verify each edited section from the Edit-return content:
**).| but doesn't end |).On failure: ⚠️ Sector note may contain a partial edit in [section]. Snapshot available: [[_Archive/Snapshots/...]]. Review manually or /rollback.
Re-reads are only required if an Edit-return is genuinely ambiguous (e.g., the replacement context was truncated in the tool's response); in that case, issue a targeted section Read — not a full-file Read.
For each sector touched:
(path, snapshot_path) to tier_a_sector_snapshots.path to tier_b_sector_appends.Checkpoint B fires once at end of Step 4 — one Edit populating ## Sector notes touched.
Consult skill_origin_theses. A macro's "affected thesis set" is every changed thesis that wikilinks this macro (via _graph.md Macro reverse index OR Step 1.3 body-grep fallback).
Gate rule: skip Step 5 entirely for this macro ONLY IF BOTH conditions hold:
skill_origin_theses, ANDmacro_targets_per_research_note (computed at Step 1.2.5).If either condition fails, proceed to Step 5.0. Log (skip path): ℹ️ Skipped macro update [Macro Note] — all [N] affected thesis changes are skill-origin AND no co-changed research note resolves to this macro per Step 1.2.5.
Condition (ii) is NOT satisfied by thesis-target idempotency. See §1.11 for the 2026-04-22 failure case that motivated the explicit Step 1.2.5 / lookup-based gate separation.
Mixed set: proceed normally.
Before snapshotting:
## Log, scan for ### YYYY-MM-DD entry with entry_date == today referencing the source.
ℹ️ Skipped macro update [Macro Note] — Log already contains today's entry for [research-note].## Log section → fall through to step 3.## Log): if wikilink present AND mtime > .last_sync, skip. Log: ℹ️ Skipped macro update [Macro Note] — wikilink present and mtime > .last_sync (no Log section to confirm); likely propagated within current sync window. Forced re-propagation: /sync all or manual edit.See §1.10 for mtime fallback trade-offs.
If edits modify existing analytical text (scenario analysis, probability weightings, trading/allocation implications), snapshot ALL Tier A macro notes in one Bash block using background shell jobs + wait:
# For each macro in tier_a_macro_snapshots, fire cp in background:
cp "Macro & Technology/Note A.md" "_Archive/Snapshots/Note A (pre-sync YYYY-MM-DD-HHMMSS).md" &
cp "Macro & Technology/Note B.md" "_Archive/Snapshots/Note B (pre-sync YYYY-MM-DD-HHMMSS).md" &
wait
Then issue all frontmatter-addition Edits as a single parallel tool-call batch.
Add frontmatter to each snapshot:
snapshot_of: "[[Macro & Technology/Note Name]]"
snapshot_date: YYYY-MM-DD
snapshot_trigger: sync
snapshot_batch: sync-YYYY-MM-DD-HHMMSS
Skip snapshot if only adding wikilinks.
For each macro touched:
(path, snapshot_path) to tier_a_macro_snapshots.path to tier_b_macro_appends.Checkpoint C fires once at end of Step 5.
Follow .claude/skills/_shared/hot-md-contract.md for all writes — compression, per-section budgets, truncation-marker avoidance, cap handling defined there.
If _hot.md does not exist, create with sections: ## Active Research Thread, ## Latest Sync, ## Sync Archive, ## Recent Conviction Changes, ## Open Questions, ## Portfolio Snapshot. Frontmatter: date: YYYY-MM-DD, tags: [meta, hot-cache].
All-or-nothing: either every planned section write lands, or none of them do. Prior spec wrote sections 1-6 sequentially then checked the hard cap — so a run hitting the hard cap left _hot.md with sections 1-3 updated and sections 4-6 at their prior (stale) state.
The M5-compliant order is: read → stage in memory → pre-check cap → compress if needed → pre-check again → commit writes OR abort before any write.
_hot.md into memory (non-destructive)Read current file. Parse sections. Note the current word count of each of the 6 canonical sections.
For each planned write, compute the projected section text in memory (do not Edit yet):
date: in frontmatter and heading.*Previous:* entry (date + one-phrase summary); max 5 *Previous:* lines; drop oldest. See hot-md-contract.md §5.- **[TICKER]**: [what changed] — [conviction impact]
5a. Mark resolved Outstanding Questions: if any resolved on theses during Step 3d, stage the update. If conviction triggers or drift flagged in Step 3e, stage the reassessment question.
5b. Auto-resolve /catalyst no-catalyst questions (§7.1). Scan staged ## Open Questions for the /catalyst-emitted pattern:
- [[TICKER]]: what catalyst in next 90 days could reignite this thesis? (flagged by /catalyst YYYY-MM-DD)
For each match, resolve if the linked thesis now has a forward catalyst. Method:
Theses/TICKER - *.md's ## Catalysts section.YYYY-MM-DD, Q3 2026, October 2026, next [N] weeks/months).>= today, thesis has a forward catalyst → stage removal (or strikethrough) of the Open Question entry.
_hot.md already shows ~~strikethrough~~ → Resolved pattern): - ~~[[TICKER]]: what catalyst...~~ → Resolved YYYY-MM-DD: forward catalyst now exists ([summary]).Skip safety: past-dated catalysts only → question stays. Missing/unparseable ## Catalysts → ℹ️ Cannot auto-resolve /catalyst question for [TICKER] — Catalysts section missing or unparseable. (emit to Step 8 report; do NOT stage a write that can't be computed cleanly).
Report resolved tickers in Step 8 under /catalyst Open Questions auto-resolved.
Compute projected_total = sum(staged section word counts) + frontmatter + heading overhead.
projected_total ≤ soft_cap (4000): proceed to 6.4 direct commit.soft_cap < projected_total ≤ hard_cap (5000): proceed to 6.3 staged compression.projected_total > hard_cap (5000): attempt 6.3 staged compression. If still over hard cap after full compression trigger order, ABORT all _hot.md writes for this run (§6.5) before any Edit lands. /hot.md retains its pre-run state. Report the abort in Step 8.Apply compression trigger order from _shared/hot-md-contract.md to the staged section buffers:
*Previous:* line from staged Active Research Thread.## Outstanding Questions).§Log + body).After each step, recompute projected_total. Stop when under soft cap. If all six exhausted and projected_total > hard_cap, return to 6.2's abort path.
Only reached when projected_total ≤ hard_cap. The prior six-Edits form serialized the all-or-nothing commit into six dependent tool calls, any of which could fail mid-sequence and leave partial state. Per .claude/skills/_shared/hot-md-contract.md §5, section ORDER inside _hot.md is not load-bearing (no consumer reads sections in sequence). The commit therefore compresses to one composite Edit that replaces the entire post-frontmatter region with the staged content — a single tool-call round-trip instead of six, and true atomicity (the Edit either lands whole or does not land at all).
Procedure:
date: and main heading.## Active Research Thread, ## Latest Sync, ## Sync Archive, ## Recent Conviction Changes, ## Open Questions, ## Portfolio Snapshot (mirrors the _hot.md schema from CLAUDE.md Rule #9).---\n plus the current post-frontmatter content:
old_string = the closing ---\n followed by the CURRENT post-frontmatter content (captured at Step 6.0's Read).new_string = the closing ---\n followed by the STAGED post-frontmatter content.If the composite Edit fails (rare — pre-checks at 6.2 should make this the pathological case only, and anchor uniqueness is guaranteed because the closing ---\n appears exactly once in a well-formed file), report the failure and do NOT fall back to sequential Edits. _hot.md retains pre-run state; primary sync propagation in Steps 3-5 is unaffected. /lint #35 catches any post-run schema drift; /lint #42 catches truncation markers.
Frontmatter-only update fast path: if date: is the only field that changed between the read and the staged output (body sections all byte-identical after compression), skip the composite region Edit and issue a single targeted frontmatter Edit on the date: line. One-line Edits are cheaper than whole-region replacements when the body didn't actually change.
If 6.2/6.3 aborted before any write:
⚠️ _hot.md hard cap (5000 words) exceeded after all compression triggers — no _hot.md writes committed this /sync run.
Current: [X] words. Projected with planned writes: [Y] words. Over cap by: [Z] words.
Pre-run _hot.md state preserved. Resolution options:
(a) Run /lint #42 to surface accumulated truncation/compression markers.
(b) Manually compress _hot.md (drop oldest Sync Archive entries, *Previous:* lines).
(c) Accept the session-context degradation for this run; primary sync propagation already succeeded.
Primary sync operation (Steps 3-5) remains valid; only _hot.md is skipped. Preserves the contract's principle: session-context degradation is tolerable; primary operation correctness is not.
Runs after ALL content writes (Steps 3-6) complete, before Step 7.5.
.sync_all_fresh writes BEFORE .last_sync touch. If either fails, this ordering preserves correctness:
/sync re-reads everything (safe — wasted work but no missed rebuild)./graph still forces full rebuild from marker; next /sync re-reads (safe — doubly covered)./graph last skips full rebuild → "no delta" theses unreconciled in _graph.md until next /sync all./sync all only (writes FIRST per C4 ordering)Skip for /sync (default) and /sync TICKER. For /sync all:
touch .sync_all_fresh
/graph reads this marker at Watermark Resolution and forces full rebuild regardless of mode, then deletes the marker after a successful write. See §5.4.
/sync / /sync all: touch .last_sync. If prior marker write failed, watermark stays unset and next run re-processes.
touch .last_sync
/sync TICKER: do NOT touch (§6.1). Report Watermark: untouched (ticker-scoped mode preserves baseline for next default /sync).
.last_sync absent, create epoch placeholder:
touch -t 197001010000 .last_sync
Watermark: epoch placeholder created.Mandatory for every mode when Step 2.9 skeleton was written. Skip only on pure no-op.
Reuse the batch ID captured at Step 2.9a. Path:
_Archive/Snapshots/_sync-manifest (sync-YYYY-MM-DD-HHMMSS).md
Edit manifest frontmatter:
status: in-progress → status: completedcompleted_date: YYYY-MM-DDExpected frontmatter post-flip:
---
type: sync-manifest
batch: sync-YYYY-MM-DD-HHMMSS
mode: default | all | ticker-scoped
status: completed
date: YYYY-MM-DD
completed_date: YYYY-MM-DD
---
Phase checkpoints A/B/C have already populated the body. No body Edit at this step.
Inspect the 7.5b Edit tool call's return value. The Edit tool reports success iff the replacement landed; the return value shows the post-edit frontmatter snippet. Confirm from the Edit-return content:
status: completed presentstatus: in-progress absentcompleted_date: present and equals todayOn success: proceed to Step 8.
On verification failure (Edit-return indicates the replacement did not produce the expected frontmatter, or the Edit tool returned an error): do NOT retry aggressively. Report ⚠️ Sync manifest status flip failed — manifest at [path] remains status: in-progress despite successful sync completion. /lint #41 will flag this as Important until manually resolved. Manual fix: edit the manifest and replace status: in-progress with status: completed (add completed_date: today). Continue to Step 8.
Throughout Steps 3-5, maintain in-memory accumulators flushed at phase checkpoints:
| Accumulator | Populated when | Flushed at |
|---|---|---|
tier_a_thesis_snapshots: [(path, snapshot_path)] | Step 3c snapshot | Checkpoint A |
tier_b_thesis_log_appends: [(path, log_entry_text, reason)] | Step 3f without snapshot | Checkpoint A |
source_research_notes: [path] | Step 3 starts on a research note | Checkpoint A (dedup) |
tier_a_sector_snapshots: [(path, snapshot_path)] | Step 4a snapshot | Checkpoint B |
tier_b_sector_appends: [(path)] | Step 4b wikilink-only | Checkpoint B |
tier_a_macro_snapshots: [(path, snapshot_path)] | Step 5a snapshot | Checkpoint C |
tier_b_macro_appends: [(path)] | Step 5b wikilink-only | Checkpoint C |
If the run completes with all accumulators empty (pure no-op reached Step 3 but no edits landed): skip Step 7.5 entirely. In practice this is rare since Step 1 catches the pure no-op earlier.
The manifest is a non-snapshot artifact (type: sync-manifest, no snapshot_date:):
/clean Step 2a's non-snapshot guard skips it for snapshot-age-based cleanup./lint #41 distinguishes three states: status: in-progress → Important; status: completed → ages per 90/180 day tiers; missing → no-op./lint #41 ages it out.Produce a compact report. Emit only lines with non-empty data. Do NOT include "If none, omit this line" scaffold; if the field is empty, skip the line entirely. The LLM should inspect each field before deciding to write a line, not write a template and then prune.
Always emitted:
default | all | ticker-scopedtouched (default/all) | untouched (ticker-scoped mode preserves baseline for next default /sync) | epoch placeholder created→ Run /graph last to update the dependency map. Recommended after every /sync.Conditionally emitted — include only if data is non-empty:
| Field | Emit when |
|---|---|
Snapshots created | at least one Tier A snapshot (theses/sectors/macros) — list paths |
Thesis updates | at least one thesis touched — one bullet per thesis: [TICKER]: [sections] — [conviction impact]. Snapshot: [[...]] or N/A |
Sector note updates | at least one sector touched — one-line summary each |
Macro note updates | at least one macro touched — one-line summary each |
Deduplication skips | at least one Step 1.7 wikilink-presence skip |
Sector skill-origin skips | at least one Step 4.-1 skip |
Macro skill-origin skips | at least one Step 5.-1 skip |
Skill-origin classified theses | at least one thesis classified as skill-origin in Step 2.5 — list each as [TICKER]: most-recent Log prefix "[prefix]" (sector/macro propagation skipped). The Step 2.5 Body-change override now actively catches the manual-edit-without-Log-entry case (CLAUDE.md Workflow Rule #6): if a skill-origin thesis's body changed outside ## Log since last sync, the user was prompted mid-run and (on Y) it was reclassified research-driven — so a thesis remaining in this skill-origin list either had only Log-section changes (correct skip) or the user chose n at the prompt (or the no-git fallback found no prior snapshot to compare). For any surprise here, append a Manual edit: ... Log entry and re-run /sync. Also surface Body-change overrides prompted: [tickers, Y/n each] when the override fired. |
Log-history backfill skips | at least one Case 2a skip (Step 1.7) |
Sector idempotency skips | at least one Step 4.0 skip |
propagated_to: skips | at least one Case 2b skip |
Terminal-skip notes | at least one Case 2c (empty propagated_to: []) |
Augmented targets | at least one augmented ticker |
propagated_to: updates | at least one research note had its frontmatter updated |
propagated_to: update skipped | at least one research note had an append-failure blocking atomic update |
Closed-status skips | Step 1.5 caught a status: closed thesis still in Theses/ |
Missing-file skips | graph-referenced thesis missing from disk |
Unresolved research notes | no propagation target found for at least one note |
Conviction triggers hit | Step 3e ⚡ fired on at least one thesis |
Conviction drift flags raised | Step 3e ⚠️/📈 fired |
/catalyst Open Questions auto-resolved | Step 6 #5b resolved at least one |
New connections discovered | at least one surprising link noted during analysis |
Final line (always): → Run /graph last to update the dependency map. Recommended after every /sync.