Repository navigation
Conversation
Repackage claude-vault-capture as a Claude Code marketplace plugin under skills/claude-vault-capture/. It auto-captures Claude Code sessions into an Obsidian vault on SessionEnd (scrub secrets → summarize with Sonnet → write a curated note to Inbox/auto/) and ships the /vault-save skill for on-demand exports to claude-docs/. Adapted from https://github.com/lhoupert/claude-vault-capture (MIT, © Loïc Houpert). Changes made for marketplace distribution: - plugin.json with userConfig (vault path + sensitive API/OAuth tokens) replaces install.sh's interactive prompt and capture.env. - hooks/hooks.json registers the SessionEnd hook via ${CLAUDE_PLUGIN_ROOT}, replacing the manual settings.json edit. - vault-save becomes skills/vault-save/SKILL.md; its auto-trigger lives in the skill description, so the old global CLAUDE.md injection is gone. - curate.py carries PEP 723 inline deps and the hook launches it with `uv run` (subscription mode adds claude-agent-sdk via `uv run --with`), so there is no separate `uv sync` step. A pre-built .venv is still preferred when present, so standalone/dev use and the existing test suite are unchanged. - Runtime state (dedup index, logs) moves to ${CLAUDE_PLUGIN_DATA} via CAPTURE_STATE_DIR / SCRUB_FAILURES_PATH, surviving plugin updates. - LICENSE and a provenance note are preserved; the pytest suite travels along. Register the plugin in marketplace.json and list it in the README. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
tests/test_timeout.py imports anthropic/httpx at module level, but the PR moved the runtime dep into PEP 723 metadata only, so 'uv sync && uv run pytest' (the documented flow) failed at collection. Pinning the same version as curate.py's inline block also unbreaks the standalone .venv hook path, which ran curate.py without anthropic installed. uv.lock is committed for reproducibility (dev deps are exact-pinned). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…-mode tests The uv-run branch expanded an empty WITH array under 'set -u', a fatal 'unbound variable' on bash < 4.4 — macOS ships 3.2, so every capture from a marketplace install (no .venv, API-key mode) died before launching the worker. Build the RUN array incrementally instead. Fallback credential files (~/.claude_vault_token, ~/.claude_vault_oauth_token) are now refused unless owner-only (mode *00), with a CAPTURE_TOKEN_FILE_PERMS marker logged; README documents umask/chmod 600. New tests/test_plugin_mode.py runs the hook copy under /bin/bash with a stub uv and no .venv — the exact configuration the old suite never exercised (its harness always built a .venv, so the uv branch was dead code in CI). Covers CLAUDE_PLUGIN_OPTION_* mapping, CLAUDE_PLUGIN_DATA state dir, the subscription --with arm, env-precedence, and token-file perms; 6 of 7 fail against the unfixed script. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ject JSON path render_frontmatter interpolated scalars unquoted, so the house-style 'Decision: …' titles produced frontmatter yaml.safe_load (and Obsidian) reject; every string scalar is now json.dumps-quoted, which is valid, inert YAML. Model-supplied tags/type were the one output-to-disk channel that bypassed sanitization: tags are now scrubbed and slug-coerced ([a-z0-9-], max 40 chars, 10 tags), type is allowlisted. Valid-but-non- object model JSON (quoted string, list, number) now routes down the malformed_json path with usage attached instead of an AttributeError that dropped token accounting. conftest.parse_frontmatter now uses yaml.safe_load — the old partition- on-colon parser is what masked the invalid-YAML bug; the suite now fails exactly when Obsidian would. New round-trip and sanitizer tests cover colon titles, hostile tags/type, and the non-object JSON cases. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
env_var was ^-anchored, so 'export KEY=…' lines and assignments on a message's first line (which curate.py prefixes with '[USER]: ') passed through unredacted, and quoted values leaked everything past the first space. The rule now matches after whitespace/quote/paren/colon, accepts an export/set/env prefix, and consumes quoted values whole; lowercase keys still never match. token_prefix's sk- charset stopped at the first dash, redacting only the public 'sk-proj' prefix of modern OpenAI keys while the key body stayed in clear; it now consumes _/- and covers github_pat_ fine-grained PATs. A new aws_secret rule catches the lowercase ~/.aws/credentials forms (aws_secret_access_key, aws_session_token) the uppercase-only env_var rule missed. The with-secrets fixture no longer stores a real-format PEM block: gitleaks/push-protection match the BEGIN/END markers themselves, so load_fixture() assembles them at load time instead. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…stale references The repo README's manual-install/symlink instructions silently fail for this entry (nested SKILL.md, hook only registers under /plugin install), so the entry is now marked code-bearing/plugin-only with a caveat in the manual-install section, and the intro acknowledges the marketplace hosts both plain-language skills and code-bearing plugins. plugin.json's repository now points at this marketplace (where the packaged code lives); homepage keeps the upstream attribution link. curate.py comments no longer reference install.sh flows or files that don't ship here. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Brings the vendored copy up to developmentseed/claude-vault-capture@c03ccda (the repo moved from lhoupert/; the old URL now redirects). Ports five upstream commits that landed after the v0.2.0 the plugin was vendored from: configurable CAPTURE_TIMEOUT_SECONDS, tool activity surfaced to the curator, the W29/W30 no-capture fixes (JSON salvage from prose-wrapped replies, subscription error_max_turns, unknown-usage propagation), and the simplification refactor. Test count goes 189 -> 223. Three merge conflicts resolved rather than taken wholesale: - scrub_rules.py: upstream replaced the {sentinel, replace_value_only, replace_group} rule schema with a single 'replacement' regex template. The review fixes' rules were re-expressed in it — env_var keeps its widened boundary (export/set/env prefixes, first-line '[USER]: ' spans, quoted values) with the pre-value span captured as \g<pre>, and aws_secret gains a 'replacement'. Without this the new engine's _compile_rules would have caught the missing key and SILENTLY SKIPPED the AWS rule. - curate.py: upstream's JSON salvage and the review's non-dict handling are complementary; merged so a hard parse failure and valid-JSON-of-the-wrong- type both attempt salvage before raising malformed_json with usage attached. A literal null payload now retries like a raw "null" instead of crashing. - session-end-capture.sh: upstream's timestamp hoist kept; its git-based deploy-drift guard scoped to standalone mode, since an installed plugin can't drift that way ( resolves to the marketplace clone) and it would spend forks on the close path to log an irrelevant SHA. Verified it still fires standalone and stays quiet under CLAUDE_PLUGIN_ROOT. Version 0.2.0 -> 0.3.0 (upstream's own tag is still v0.2.0; these commits are unreleased there). New knobs documented in the plugin README. Provenance URLs follow the repo move; the MIT copyright (© Loïc Houpert) is unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…n mode Two gaps a cross-check of the port surfaced: The headline ported feature was unreachable. CAPTURE_TIMEOUT_SECONDS is the fix for the upstream failure that motivated it (subscription-mode timeouts on large transcripts) — and surfacing tool activity makes every transcript bigger — but plugin config only reaches curate.py through userConfig, which had no entry for it. Adds timeout_seconds and maps it in the hook. The mapping validates: curate.py reads the var with a *string* default, so an empty or non-numeric value would be a ValueError at import — a silent no-capture — rather than a fallback to 30. Junk is logged and dropped. Deploy identity is now logged in both modes instead of being suppressed for plugin installs. Scoping the git guard to standalone (previous commit) threw away the W29 lesson — you could no longer tell from hooks.log which version produced a capture, in the mode where that is hardest to reconstruct. Plugin installs now log the plugin version from plugin.json; git stays standalone-only because "git -C" walks up and would describe whatever repo encloses the plugin, reporting false STALE_DEPLOYs against an unrelated origin/main. Also hoists the mkdir/NOW pair above all config handling so every logging path has its destination, and documents the new knobs plus the tool-activity change in the plugin README. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…urates it Brings the vendored copy to developmentseed/claude-vault-capture@7577892, one commit past the previous port. Upstream's root-cause fix for every malformed_json in its archive: the prompt was an unterminated chat log, so the model wrote the conversation's next turn instead of an artifact. Ports all four layers — the _TRANSCRIPT_TAIL terminator appended by _invoke_model (so both transports get it), _salvage_artifact's offset-scanning replacement for the old outermost-brace window, the shared PATH_A_RESAMPLES budget that now retries an unparseable reply once, and the held `unparsed` error so a trailing null can't relabel a malformed reply in log.md. Suite 227 -> 237. One conflict, in the same block the last port resolved. Upstream rewrote _call_path_a's failure handling; the review's non-dict guard is still needed because upstream does not have it — a reply that parses to "null", [1,2] or 42 never enters the except branch, so it reaches data.update() and dies as an AttributeError, losing the usage accounting. Rather than keep it as a separate raise-immediately path, the wrong-type class is now folded into upstream's recovery route: salvage, then one resample, then malformed_json with both attempts billed. That is upstream's own argument for retrying prose replies — a wrong-shaped reply is non-deterministic too — and it keeps one failure path instead of two. A literal parsed null still resamples as a null, matching upstream. Verified directly: all three wrong-type shapes raise JSONDecodeError with usage after 2 attempts, no AttributeError. _salvage_object (added by the previous port) is deleted — upstream's _salvage_artifact supersedes it and is strictly better: it scans every brace offset instead of taking a first-to-last window that prose could widen, and it requires the title/type/body contract, which closes a junk-write hole where an unshaped dict would land in Inbox/ as an empty "untitled" note. Version stays 0.3.0: that release is still unmerged, so this folds into it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ber gaps, unlogged losses
A 26-agent review with adversarial verification confirmed 21 findings against
this branch (0 refuted). The security-critical one and the defects that lose
data or corrupt notes are fixed here; each fix was reproduced before and
verified after. Suite 237 -> 255.
CRITICAL — secrets survived the scrubber via truncation. render_transcript caps
tool blocks (_BASH_CMD_CAP, _EDIT_DIFF_CAP, _ERROR_CAP, CAPTURE_SUCCESS_HEAD_CHARS)
and the caps ran BEFORE the scrub, so a cut secret matched no rule: private_key
needs its closing -----END----- and AIza…{35}/AKIA…{16} need their full body.
A PEM pasted into a failing command reached the model and the note in clear
(reproduced on [OUT], [TOOL] Write, and [ERROR]). Scrubbing now happens per block
before each cap, via _scrub_cap; render_transcript takes an optional dict to
collect those counts so `redactions:` still totals them.
Scrubber, two rules:
- env_var never matched a bare keyword name. A mandatory leading [A-Z] sat
outside the alternation, so the keyword could never start at offset 0 —
PASSWORD=, TOKEN=, SECRET=, PASS=, KEY_ID= all passed through in clear while
the frontmatter reported 0 redactions. These are the commonest .env forms.
- the widened sk- alternative corrupted ordinary prose. Adding `-` to the body
charset let it run across segments: "risk-averse-approach" became
"ri<redacted:token_prefix>" in the archived note. Now word-boundary anchored,
no `-` in the body, 16-char minimum. Both directions are now tested.
Unlogged session losses. Only the salvage path enforced the title/type/body
contract, so a directly-parsed reply with a non-string body (or title, or a
non-list tags/source_links) raised inside the write path — after the model call
was paid for and before append_log ran, leaving no log.md row at all, which is
the failure class the module comments call out as invisible to the no-capture
alarm. The contract is now shared by both paths via _is_artifact, and the
containers are coerced. Verified: all four shapes now log rather than crash.
Hook script:
- stat now follows symlinks (-L) and probes GNU -c before BSD -f. A symlinked
token was refused with a chmod hint that could not fix it, and on Linux the
BSD-first probe wrote a multi-line stat blob into hooks.log as the mode.
- the plugin.json version sed gets `|| true`: under set -euo pipefail a missing
file made it abort the hook, skipping the capture over a cosmetic log line.
- an installed plugin now always uses `uv run`, never a .venv. Running `uv sync`
inside an installed plugin — which the README's own Tests section invited —
left a dev-only venv that shadowed the PEP 723 deps and silently broke
subscription mode. README now says to run tests from a clone.
- a missing python3 is logged (CAPTURE_NO_PYTHON3) instead of exiting silently,
and python3 is listed under Prerequisites.
Tests: the plugin-mode stub now writes atomically (rename) so the poll cannot
read a half-written record, and its timeout probe distinguishes unset from
set-and-empty — that assertion previously could not fail for the reason it
existed. One upstream assertion pinned a specific sentinel label on a value two
rules both claim; it now asserts that nothing leaks, which is the real contract.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…fig can migrate Found while migrating a real standalone install: of the five settings a typical capture.env carries, only three had a plugin equivalent. CAPTURE_EXCLUDED_COMMANDS and CAPTURE_MAX_EST_TOKENS were readable by curate.py but unreachable through plugin config, so switching to the plugin silently dropped them — the excluded workflow commands (/daily-devlog, /weekly-recap) would start being captured into the vault, and a raised ceiling (120000) would fall back to 50000, skipping the long sessions it was raised for. Neither failure is visible at switch time. Adds both as userConfig and maps them in the hook. The numeric validation added for timeout_seconds is factored into _export_int_setting and now covers max_est_tokens too: curate.py reads both with string defaults, so an empty or non-numeric value is a ValueError at import — a silent no-capture — rather than a fallback. The rejection marker is now CAPTURE_BAD_SETTING (was CAPTURE_BAD_TIMEOUT) since it covers more than the timeout. excluded_commands needs no validation: curate.py splits on "," and drops empties. Verified by running the hook with all five values set as CLAUDE_PLUGIN_OPTION_* and confirming each reaches the worker unchanged. Suite 255 -> 258. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Swaps MODEL_A to claude-sonnet-5, which covers both transports — the API-key path and the subscription path both pass this one constant through. Two changes the swap requires, neither of which the model ID alone gives you: Thinking is now explicitly disabled on the API-key call. Sonnet 5 runs adaptive thinking whenever `thinking` is omitted, where Sonnet 4.6 ran without it, and max_tokens caps thinking and the reply together. Left unset, the curation call would spend the artifact's budget on reasoning and truncate the JSON — arriving as malformed_json and a lost capture rather than an error. Curation is a one-shot extraction against a hard timeout, so thinking buys little here; the comment records adaptive-at-low-effort as the tunable if capture quality ever needs it. MAX_TOKENS_A goes 2000 -> 3000. Sonnet 5's tokenizer emits roughly 30% more tokens for the same text, so artifacts that previously fit would otherwise truncate at the same ceiling. Nothing else in the request had to change: the call passes no temperature/top_p/ top_k (rejected on Sonnet 5) and no assistant prefill (also rejected). List pricing is unchanged at $3/$15 per MTok, so the cost estimator stays correct — noted inline that introductory rates make it over-report through 2026-08-31. The chars/4 token heuristic now runs ~30% low against the new tokenizer, so the CAPTURE_MAX_EST_TOKENS ceiling admits somewhat larger transcripts than the number implies. Left deliberately as-is and documented: tightening it would silently start skipping sessions that are captured today. New tests pin the request shape — thinking explicitly disabled, no removed sampling params, and output headroom for the new tokenizer. Verified the first fails when the thinking line is removed. Suite 258 -> 261. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A newcomer-focused documentation review (agents, each finding verified
against the vendored code) returned 33 confirmed issues. The ones that
actually broke a first install:
- The Install section said Claude Code prompts for a vault path "and,
optionally, an API key. That's it." Credentials are not optional:
curate.py exits before the model call with no note AND no log row, so
the only trace is one line in hooks.log. Pro/Max users — who typically
have no ANTHROPIC_API_KEY anywhere — were the most likely victims.
The same "leave blank if you set it in your environment" advice was in
the install prompt itself, which is worse, because Claude Code strips
the environment it spawns hooks with.
- "Verify it's working" could not distinguish working from broken. Both
its checks pass on a dead install: SESSION_END_RECEIVED is written
before the worker starts, and an empty Inbox/auto/ is the correct
result for a low-signal session. It now leads with log.md (the only
file that answers the question), spells out how to resolve
${CLAUDE_PLUGIN_DATA} from a shell, and adds two tables mapping every
skip_reason and every hooks.log marker to what to do.
- Cost and privacy were unstated. A plugin that makes an unattended paid
call at every session close now says so up front, with the real figure
(~$0.12 median, measured over 287 captures), and says plainly that the
transcript — including tool output — is sent to the API and that regex
scrubbing is not a guarantee.
Also: the curation prompt advertised a fifth artifact type,
devlog-snippet, that sanitize_type() silently rewrites to "decision", so
those notes were mislabeled in the vault — prompt aligned to the four
types the code, README and manifest all agree on. /vault-save now states
that it bypasses the scrubber (it writes directly; no Python in that
path). Install prompt reordered to ask for credentials before tuning
knobs, with every default and consequence stated per field. Added a
"Turning it off" section. Fixed the repo README's symlink fallback,
which produced a skill with an unexpanded ${user_config.vault_dir}, and
flagged there that this plugin is always-on unlike its neighbours. Stale
claude-sonnet-4-6 example updated after the Sonnet 5 move.
Suite: 261 passed, 1 skipped.
…pricing A json.dump without ensure_ascii=False had rewritten marketplace.json, plugin.json and mock-responses.json: literal characters became \u escapes (including the unrelated devseed-poster entry) and inline arrays were exploded one item per line. Formatting is restored from the PR base; content is unchanged (verified by json.loads equality). The ~$0.12 median was measured on Sonnet 4.6 rows priced at $3/$15. Sonnet 5 is $2/$10 (the introductory rate became standard; the planned September increase was cancelled). Re-measured on 160 post-cutover Sonnet 5 captures: ~$0.13 median. Numeric config fields now say "positive whole number". Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ent bloat - cost estimate: $2/$10 per MTok (was Sonnet 4.6's $3/$15, a 1.5x over-report in log.md and note frontmatter); pinned by a test. - scrub: vendor-prefixed sk- keys (sk-proj-, sk-svcacct-, sk-admin-) may contain '-' anywhere in the body; the old rule left everything after the first in-body dash in clear. Bare sk- keeps the dash-free body that protects hyphenated prose. - hook: timeout_seconds / max_est_tokens reject 0 (every call would time out, or every session skip) as well as non-numeric values. - render: MultiEdit falls through to the generic tool renderer; it carries an `edits` list, so the Edit branch rendered an empty diff. - _call_path_a: same outcomes with one `malformed` variable instead of bad/unparsed/data juggling. - Comments and docstrings trimmed to the why, dropping incident history, week numbers and session IDs; a dangling "SPEC §7" reference removed; test PEM markers assembled at runtime; ruff format applied. Suite: 265 passed, 1 deselected (live). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ments - make_slug: same output in 8 lines instead of 23. - session-end-capture.sh: $NOW is always set before _read_secret_file runs, so drop its date fallback; three verbose comments cut to one line. - tests: the identical _wait_for in test_plugin_mode and test_session_end_hook moves to conftest.wait_for. Suite: 265 passed, 1 deselected (live). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Subscription mode (the production path):
- setting_sources=[]: the curation call no longer loads user settings, so
this plugin's own SessionEnd hook stops firing on the curation session
(the source of the phantom `threshold` rows in log.md). Thinking is
disabled and output is capped at MAX_TOKENS_A, matching API mode.
- Return at the ResultMessage instead of waiting for CLI teardown inside
the timeout; raise on AssistantMessage.error and on an error result with
no reply, instead of parsing "Invalid API key" text as a reply.
Scrubber:
- basic_auth_url matches any scheme and an empty user (postgres://,
redis://:pw@, mongodb+srv://); sk-or-v1-/sk-lf- keys; AWS JSON
"SecretAccessKey"/"SessionToken".
- Generic tool inputs are scrubbed per string before json.dumps, whose
escaped newlines hid KEY=value lines (MCP inputs, MultiEdit edits).
- Fewer prose false positives: `X == 3000` comparisons, "highs_and_lows"
(GitHub tokens are 36+ chars), "Bearer token".
Pipeline:
- API stop_reason refusal/max_tokens -> skip `refusal`/`truncated` with
usage, not resampled (was IndexError / two paid identical calls).
- Every exception from _call_path_a carries usage across attempts.
- Salvage searches the whole reply, not the first fenced block; a first
line of `null` is a null; empty title/body is not an artifact; types
are lowercased before the allowlist; a model `_null` key is ignored.
- Timeouts and errors are no longer indexed, so a resumed session retries.
- A failed note write, or any error in main(), still logs a log.md row.
- excluded_commands matches Claude Code's <command-name> markup.
- derive_project names the main repo from a linked worktree, even after
the worktree is removed.
- Numeric settings: junk or empty values fall back to defaults.
- Dead code removed: HOOKS_LOG, _STATE_LOCK (flock suffices),
LOG_REQUIRED_KEYS (moved into its test).
Hook: one python3 parse instead of three (a JSON-null path no longer
becomes "None"); `~` in vault_dir expanded, relative paths refused
(they resolved inside the session's project); empty int settings unset.
/vault-save: title/summary/project/tags written as quoted strings (bare
scalars broke YAML on ": "), project from --git-common-dir, empty slug
falls back to `untitled`. README: new skip reasons and markers; uninstall
deletes ${CLAUDE_PLUGIN_DATA} unless --keep-data; Templater warning.
Suite: 304 passed, 1 deselected (live). Subscription changes verified
against claude-agent-sdk 0.2.89's option handling and its bundled CLI's
flag parsing, not with a live call.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…e test
- _invoke_via_subscription returns at the ResultMessage, which left the
CLI to exit during asyncio.run's shutdown ("Loop ... that handles pid
is closed" on every capture, i.e. hooks.log noise). The stream is now
closed explicitly inside the loop, after the reply and outside the
timeout, so a slow teardown still can't discard a finished reply.
- test_live_e2e crashed before any model call: the autouse temp_vault
fixture already creates vault/Inbox/auto. mkdir(exist_ok=True).
Verified live (CAPTURE_LIVE_TESTS=1, real subscription call): passes, no
asyncio warning, no orphaned CLI, and no new SESSION_END_RECEIVED line or
log.md row from the installed plugin, so setting_sources=[] stops the
plugin's own hook firing on the curation session.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rgrft2uNyq6DfUkcbE5xMD
A set `version` pins installs until it changes, so installs taken from this branch at 0.3.0 would never pick up the fixes since 29 July. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Rgrft2uNyq6DfUkcbE5xMD
Sonnet 5.5 rejects thinking {"type": "disabled"} with a 400; its
thinking-off setting is {"type": "between_tools"}, run at effort low.
- API mode sends between_tools plus output_config effort "low".
- Subscription mode: the CLI has no --thinking between_tools, so it gets
CLAUDE_CODE_EXTRA_BODY (the thinking override) and
CLAUDE_CODE_DISABLE_THINKING (drops the clear_thinking context edit
the API rejects alongside between_tools), plus effort "low". Both env
vars are undocumented CLI internals: re-check them on SDK pin bumps.
- claude-agent-sdk 0.2.89 -> 0.2.161. The old bundled CLI did not honour
--thinking disabled either: 10 of 22 Sonnet 5 curation calls thought,
3 hit the output cap.
A/B on 20 real transcripts, subscription mode, 0.3.1 vs this change:
20/20 notes vs 19/20; median 11 s vs 20 s (p90 15 s vs 41 s); median
cost $0.106 vs $0.126 per capture; blind judges preferred Sonnet 5.5
17-1 on usefulness and 9-0 on accuracy (8 distortions, all Sonnet 5).
Live check: the curation transcript has 0 thinking blocks.
Suite: 305 passed, 1 deselected (live, passed separately).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rgrft2uNyq6DfUkcbE5xMD
- Drop standalone mode: no capture.env, no .venv interpreter, no git STALE_DEPLOY check. The plugin is installed only via /plugin and the upstream repo is being retired, so that path had no users. The deploy line now always logs the plugin version. test_session_end_hook.py is removed; its happy-path, guard, unconfigured-vault and OAuth-token-file cases move to test_plugin_mode.py. - umask 077 in the hook: hooks.log, log.md, the index and new notes hold transcript excerpts and were created world-readable (0644). - API mode keeps the SDK's retries (max_retries=2): a single 529/5xx lost the capture for good. TIMEOUT_SECONDS now bounds each attempt. - README: the threshold row describes what the gate actually counts. The documented "3 real turns / 1500 chars of your own text" would have skipped 17 of 116 sessions that produced notes (measured on the live log), so the docs change rather than the code. Suite: 302 passed, 1 deselected (live). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Rgrft2uNyq6DfUkcbE5xMD
…pend - sanitize_summary had one caller: sanitize_title now takes max_len. - make_filename was one f-string with one caller: inlined; the e2e test already checks the filename format. - append_log and _append_index share _locked_append; the explicit flush()/LOCK_UN before close is dropped (close flushes, then releases the lock). - _write_artifact passes render_frontmatter's keyword args through as **fm instead of re-declaring ten of them. Drops the 13 tests that duplicated TestSanitizeTitle or tested make_filename. Suite: 290 passed, 1 skipped (live). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
|
Claude finished @lhoupert's task in 36s —— View job ✅ No blocking issues — safe to merge. I read Scrub-before-cap ordering, the model-output allowlists (type, tags, JSON-quoted frontmatter), the
Simplify (ponytail)
💰 Estimated review cost: $0.27 · 0m35s · 6 turns |
Member
|
thanks @lhoupert for bringing this over. I read through the hook,
small one: |
Author
|
Thank you for your comments Danny! I will take the time to look into it later this week. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds the claude-vault-capture plugin to the marketplace under
skills/claude-vault-capture/. OnSessionEndit scrubs secrets from the transcript, has Claude Sonnet 5.5 curate one note (decision, runbook, gotcha or spec, or nothing for low-signal sessions) and writes it to<vault>/Inbox/auto/; it also ships the/vault-saveskill. This supersedes #5 together with the review fixes stacked on it in alukach#1, so it can be reviewed and merged from this repo. 290 tests pass; ruff and shellcheck are clean.For reviewers
Unlike the other skills here, this one runs by itself: each qualifying session costs a model call (~$0.11 median, or billed to a Claude Pro/Max plan). Sonnet 5.5 rejects
thinking: disabled, so curation usesbetween_toolsat effortlow; in subscription mode that relies on two undocumented CLI env vars,CLAUDE_CODE_EXTRA_BODYandCLAUDE_CODE_DISABLE_THINKING, to re-check when bumping theclaude-agent-sdkpin.Author attestation
AI-assisted: the fixes came from multi-agent reviews, with each finding verified against the code; reviewed via the diff and the test suite.
🤖 Generated with Claude Code