Changelog

1.29.1 — 2026-08-02

• Stops a stale Codex install from blocking a session. Codex persists a hook's resolved command, so upgrading left hooks pointing at a version directory that had been removed. The interpreter exited 2 before any trigger-tree code ran, and Codex treats a non-zero UserPromptSubmit hook as a blocking error, so the user's prompt was refused. The logger's own fail-open guarantee cannot cover a file that is not there.
• Every Codex hook command now checks the script exists and exits 0 when it does not, on POSIX and Windows alike. A stale path degrades to no telemetry instead of a broken session, which is the rule this plugin holds itself to.

1.29.0 — 2026-08-01

• Reports when a project's TT_ALWAYS_LOADED_REGEX predates the current version and misses a surface the client injects. /tt setup snapshots the bundled defaults into the project, so an install created before .claude/rules/ was added to that pattern still classifies rule files as evaluable documentation: they appear as untouched and lower the health score, with nothing to say why. Reported from a live deployment.
• Counts only .claude/agents/*.md files carrying YAML frontmatter as personas. A project-level agents/ directory is documentation in this tool's own inventory, so scanning it counted READMEs and templates as agent definitions.
• Splits persona cost. A definition's name and description are injected on every request; its body is paid only when the agent runs. Charging whole files to every request overstated the recurring cost by an order of magnitude, so estimated_tokens_per_session now covers the injected part and estimated_tokens_when_invoked the body.
• /tt insights now says how to enable agent capture when personas are defined but capture is off, instead of staying silent. tt doctor and the HTML report already did.

1.28.2 — 2026-07-29

• Treats the install location as authoritative when identifying the client. A Codex marketplace install resolves its manifest's ./skills/ to the repository's Claude skill, because the portal validator requires that path, and that skill asserts --client claude. A Codex session was therefore reported as Claude. Where the files live is observable, so it now outranks the flag in tt-open.sh, tt-stats.py, and tt-watch.py. Recorded telemetry was never affected: the client field comes from the hook manifests, not this flag.
• Explains that a runtime which blocks the first watcher launch is usually asking for approval rather than failing permanently, so the retry that succeeds no longer reads as a broken command.
• Documents Codex invocation as @trigger-tree <command>, alongside the plain-language form; both reach the same workflows.

1.28.1 — 2026-07-29

• Completes the agent-persona surfaces. 1.28.0 added the measurement and the stats payload but left the rendered surfaces silent, so /tt insights published an HTML report that said nothing about personas while the skill pointed at them.
• The HTML report gains an agent-personas section with invocations, sessions, last use, and never-invoked definitions. The live dashboard gains one line with the most used persona and the never-invoked count. The published demo report shows both, since that is what most people see first.
• tt suggestions reports never-invoked personas as an observation, never as a numbered edit: this tool proposes verified router changes only, and an unused persona needs judgment rather than an edit.
• Every surface stays silent when agent capture is off, because silence would then measure missing instrumentation rather than disuse.
• Documents the new surfaces in the dashboard guide and answers the persona question in the FAQ.

1.28.0 — 2026-07-29

• Measures subagent personas. Every definition under .claude/agents/ or agents/ puts its description in the system prompt on every request, so a persona never invoked is recurring cost in exactly the way an untriggered directive is. The stats payload now reports invocations per persona, sessions, first and last use, and which definitions have never been invoked.
• Capture is opt-in through the new TT_LOG_AGENTS, written by /tt setup and absent-means-off, so no existing install starts recording more. Only the persona name is stored; the launch payload's task prompt and description are never read.
• Never-invoked is reported only over sessions where agent capture was actually running. Otherwise the status is awaiting-capture and no cost claim is made, the same distinction the adherence report draws.
• tt doctor names persona definitions found while capture is off, so silence cannot be mistaken for an unused persona.
• The Claude Code tool name and payload field were verified against real transcripts before implementation rather than assumed. Codex routes the same tool names, but whether it surfaces a subagent launch to hooks is unverified and documented as such instead of silently degrading.

1.27.2 — 2026-07-28

• Completes the issue #21 fix for Codex. Codex exports no session id, so the current-session liveness check never reaches its strict branch there and a Codex user in a Claude-only project still saw a clean pass. Every liveness line now names the clients that have actually recorded, on the recency branch Codex reaches as well as the strict one, so one client's telemetry cannot read as another client's hooks working.
• Capture was never affected on either client: hook events take the session id from the payload on stdin, so Codex and Claude telemetry has always been attributed correctly.

1.27.1 — 2026-07-28

• Fixes tt doctor's current-session liveness check, which could never run under Claude Code (issue #21, reported by @BartWaardenburg). It read CLAUDE_SESSION_ID, which Claude Code substitutes into hook command strings but never exports, so the strict branch was unreachable and every run fell through to a lenient check that passes on any recent event from any client. A project carrying only Codex telemetry reported healthy Claude hooks.
• Session identity now resolves through one shared helper that prefers CLAUDE_CODE_SESSION_ID and keeps CLAUDE_SESSION_ID and TT_SESSION_ID as fallbacks. The same variable was used for external, shell-capture, and note events, which lost their session attribution the same way. Hook events read the id from the payload on stdin and were never affected.
• Names the clients that have recorded events when the running client's session is absent, so one client's telemetry cannot read as another's hooks working.
• Scrubs ambient session ids in the test suite. The previous tests set the variable production read, so they passed either way and kept the mismatch invisible; they now assert the variable Claude Code actually exports.

1.27.0 — 2026-07-27

• Adds an optional reason to unobservable probes: subjective-condition, requires-diff, requires-judgment, or outside-capture. The unobservable count previously merged two unrelated findings. A directive with no objectively testable condition is unlikely to fire for the model either, since it cannot recognise when the rule applies any more reliably than a probe can, and that is advice the author can act on. A directive needing the diff is a boundary of this tool and says nothing about the rule.
• Reports the two separately, adding no_testable_condition and unobservable_reasons to the stats payload, and prints a line about untestable conditions only when there are any.
• The field is optional, so manifests written before it validate and report exactly as they did.

1.26.1 — 2026-07-27

• Counts every file reached through a CLAUDE.md @import chain toward the always-loaded cost estimate, including files outside the watched documentation tree. An imported memory or notes file the agent maintains itself was previously invisible, so the cost headline understated what a session is actually billed for.
• Imports resolve only to real regular files inside the project. Symlinks, parent-directory escapes, absolute paths, ~ paths, and URLs are ignored, and the walk is bounded.
• Adds always_loaded_injected to the stats payload. always_loaded_imports, always_loaded_inventory, and every untouched/heat calculation keep their existing inventory-scoped meaning, so this is additive only.

1.26.0 — 2026-07-27

• Adds tt instructions --selftest, which proves each probe can fire by constructing evidence in memory and running the probe against it. A directive sitting at zero was previously ambiguous between "the rule never applied" and "the probe is written so nothing could ever match it"; the two are now distinguishable without running an agent, reading history, or spending tokens.
• Reports a route_followed probe whose declared path does not exist as unsatisfiable. Reads of a renamed or deleted doc can never be recorded, so the directive would otherwise sit at zero forever while looking merely unlucky.
• tt doctor now fails when any probe can never fire, ahead of capture warnings, because that is broken authoring rather than a missing setting.
• The self-test exits 1 on an unreachable or unsatisfiable probe so CI can hold a manifest honest. It proves reachability only and says so: a command pattern is never tested against the commands you actually run.

1.25.2 — 2026-07-27

• Binds recorded adherence evidence to a probe fingerprint (declared topics and command patterns) instead of the whole manifest. Editing, rewording, or moving a line of CLAUDE.md and refreshing the hashes now keeps every prior opportunity, so the before/after trend the feature exists to show survives the edit that motivates it. Changing a probe's topics or patterns still discards evidence captured under the previous semantics.
• Separates "the rule never applied" from "nothing was watching". A directive is only never-triggered over sessions where its capture was actually running; otherwise it is the new awaiting-capture, which carries no cost implication. The always-loaded cost headline uses that scoped denominator rather than the total session count.
• Reports a path_avoided rule that nobody breached as no-violations-observed rather than never-triggered, so a rule that is working is no longer counted as unused context in the cost line.
• Makes tt instructions --check state how many observable directives it actually measured, and adds --min-measured N so CI can refuse to pass on an empty evidence set.

1.25.1 — 2026-07-27

• Fixes the Codex upload manifest to resolve its skills from the packaged skills/ root. The deterministic ZIP builder already placed the skill there, but the v1.25.0 manifest still pointed at the repository-only codex-skills/ source directory, so Codex rejected the upload.
• Polishes the marketing site's comparison and health sections with consistent inset spacing, contained borders, and cleaner mobile stacking.

1.25.0 — 2026-07-27

• Adds deterministic instruction-adherence measurement: a hash-bound, schema-validated directive manifest, six observable probe types plus explicit unobservable probes, per-session evidence, confidence and maturity labels, compaction-aware uncertainty, weekly trends, and tt instructions commands for initialization, inspection, explanation, and CI checks.
• Extends local hook capture with bounded topic labels, successful in-project edit paths, classified command-pattern IDs, skill use, and observable commit boundaries. Capture gaps are reported as capture-disabled or unobserved, never silently counted as instruction violations.
• Keeps the new telemetry privacy-preserving and opt-in per project: topics and command capture default to off, edit capture defaults to matching nothing, setup explains and enables bounded topics, classified commands, and a conservative edit scope, and raw prompts, diffs, command output, and external paths are never required for adherence metrics.
• Documents the manifest and probe contract, observability limits, configuration and privacy boundaries, injected instruction imports, dashboard/report interpretation, and the distinction between static instruction validation and local evidence-based adherence.

1.24.0 — 2026-07-23

Both findings below were reported by @BartWaardenburg — thank you.

• The structural inventory follows git's view of the repository (#16): a file excluded by .gitignore can no longer surface as a router member, orphan, or missing entry point in tt-stats or tt gate. The scope scanner and the structure inventory now share one git-visible source of truth; the filesystem walk remains the fallback outside git repositories.
• Watched surfaces behind a directory symlink are named instead of silently missing (#15): the stats payload gains unfollowed_surfaces (path, target, reason), the health grade states its scope explicitly (evaluable docs), a health driver counts unfollowed surfaces, and tt doctor warns with the paths. Symlinks are still never followed — bytes outside the repository stay out of the deterministic score by design; the boundary is documented in the heat-model guide.

1.23.2 — 2026-07-23

• Codex tag pinning now pins the installed bytes (#6 reopen, verified by @BartWaardenburg): the Codex marketplace manifest hardcoded its plugin source to the repository's main ref, so a marketplace added with --ref vX.Y.Z still installed whatever main contained. The source is now relative (./), which makes a pinned marketplace checkout install its own bytes — verified against Codex CLI 0.145.0 with a checkout whose version deliberately differed from main. A release-integrity guard now fails any change that reintroduces a hardcoded ref. Pinning works for tags from v1.23.2 onward; earlier tags carry the old manifest and still resolve main.

1.23.1 — 2026-07-23

• The dashboard launcher fails gracefully in GUI-less runtimes (#14): sandboxed application runtimes such as Codex Desktop have no GUI session for AppleScript, which made osascript die mid-script under set -e without a word. Every launch route now detects its own failure, cleans up the generated temporary launcher, and prints the exact tt-watch.py command to paste into a second terminal. Running from a real terminal is unaffected.
• The Codex directory subtitle in the plugin manifest respects the OpenAI portal's 30-character limit.

1.23.0 — 2026-07-23

• User-wide configuration layer (#13, requested by @BartWaardenburg): every key now resolves as bundled plugin default → user-wide ~/.trigger-tree/config.sh (location overridable via TT_USER_CONFIG) → project override, so one person can set a privacy default such as TT_LOG_PROMPTS='off' for every repository before any project has run setup — without patching the installed plugin snapshot. The project's own choice always wins, and tt doctor reports the effective prompt mode with the layer that selected it.
• Config readers anchor assignments to the start of a line everywhere, so a commented example in any config layer can never be parsed as a value.
• Copy-paste CI examples: examples/github-actions.yml and examples/gitlab-ci.yml, linked from the README and the CI-gate guide.
• The configuration reference now states the actual pre-setup prompt default (hash); it still described the old truncate default.

1.22.0 — 2026-07-23

• GitLab CI support for the gate: tt gate --code-quality PATH writes a deterministic CodeClimate issue list — the format GitLab renders as a Code Quality widget on merge requests — with a stable fingerprint per rule and path so pipelines announce only new findings. The CI-gate guide documents the .gitlab-ci.yml job, the artifact wiring, and the badge route; the gate itself was already a plain pip-installable CLI, so no wrapper is needed on GitLab.

1.21.3 — 2026-07-23

• Corrects the GitHub release body footer: the Codex install is generated as codex plugin marketplace add Hedde/trigger_tree --ref vX.Y.Z plus codex plugin add trigger-tree@trigger-tree, pinned to the release's own tag — codex plugin install does not exist (#12, reported by @BartWaardenburg). The seventeen published release bodies carrying the dead command (v1.13.0–v1.21.2) were corrected in place, and the docs-currency guard now requires every install command in the footer template to exist verbatim in the README so the surfaces cannot drift apart again.

1.21.2 — 2026-07-23

• Documents Codex hook trust: after an install or upgrade Codex silently skips untrusted hooks, and non-interactive codex exec runs never persist trust — the quick start, FAQ, and website now say to review and trust the four hooks in the TUI, and to repeat the review after an upgrade that changes a hook (#11, reported by @BartWaardenburg).
• tt doctor reads the persisted Codex trust state ($CODEX_HOME/config.toml) whenever a Codex install carries the plugin: untrusted or partially trusted hooks warn with the exact remedy instead of a generic missing-liveness hint. The trusted hashes themselves are Codex-internal, so doctor states what it can verify and what it cannot (#11).
• SARIF output reports the real tool version for pip installs by falling back to the package metadata when the plugin manifest is not in the wheel.
• The README quick start renders as per-client command blocks instead of a two-column table that misaligned on narrow viewports.

1.21.1 — 2026-07-23

• Parses git-listed paths NUL-separated and decodes them as UTF-8 explicitly, so markdown files with non-ASCII names are counted instead of silently dropped by git's quoting — or mangled by the Windows locale codepage (#8 follow-up).
• Stamps newly written events without an attributed client as client: "unknown" explicitly — absence now only means pre-1.21 history (#9 follow-up).
• Corrects the FAQ: Codex does support version pinning via codex plugin marketplace add owner/repo --ref vX.Y.Z; the earlier claim that installs always follow the default branch was outdated (#6 follow-up).
• Removes contradictory privacy sentences the v1.21.0 rewrite left in the FAQ and SECURITY.md; every surface now states the same rule: hash before a project's consent, truncate only after setup chose it (#7 follow-up).

1.21.0 — 2026-07-23

All five findings below were reported by @BartWaardenburg — thank you.

• Restores hash as the prompt-logging fallback before a project's setup: plugin installs are user-wide, so no repository records prompt text without its own explicit choice (#7). Setup still recommends truncate interactively, and the privacy documentation states the boundary precisely.
• Scans documentation the way git sees it — tracked plus untracked-but-not-ignored markdown via git ls-files, filesystem walk as fallback: .agents/ and .codex/ are no longer hard-skipped, ignored scratch trees cannot crowd out real docs, and a capped scan reports itself as incomplete (#8).
• Stamps every telemetry event with its originating client (claude, codex, external), carried from the hook adapters through shell capture and ingest; stats expose a per-client breakdown and treat missing values as unknown (#9).
• Supports Python 3.14 across CI, doctor, and packaging, and doctor reports the interpreter path so hook and diagnostic environments can be compared (#10).
• Documents the real Codex install command (codex plugin add trigger-tree@trigger-tree) and the current marketplace ref limitation (#6).

1.20.0 — 2026-07-23

• Adds TT_SCOPE_IGNORE: comma-separated globs that acknowledge intentionally unwatched markdown (issue templates, changelogs). Acknowledged files leave the gate findings, the SARIF report, and the watch-scope denominator through one committed, reviewable config line; a path the watch regex matches is never ignored. The gate, doctor, and their shared scanner all honor it, which also auto-resolves the code-scanning alerts this repository had filed against its own human-only files.
• Watches this repository's Codex workflow contract and links it from the docs router; with the acknowledgments in place the dogfood discoverability score is an honest 100.
• Anchors config parsing to line starts so commented example values can never be mistaken for assignments.

1.19.2 — 2026-07-23

• Ships the inline README badge row in a release, because the GitHub Marketplace listing renders the README of the latest published release rather than the default branch — the fix on main alone never reached the listing.

1.19.1 — 2026-07-23

• Shortens the action description to fit the GitHub Marketplace 125-character limit, unblocking the Marketplace listing.

1.19.0 — 2026-07-23

• Exports gate findings as deterministic SARIF 2.1.0 (--sarif, action input sarif): one rule per category with warning/note levels, file locations for per-file code-scanning annotations, and score/components/verdict in the run properties; this repository uploads its own report to GitHub code scanning.
• Stops the test suite from appending fixture reports to the real CI job summary: the hermetic-environment scrub now also covers GITHUB_STEP_SUMMARY, so the run page shows exactly one gate verdict, now prefixed ✅/❌.
• Bumps the pinned action reference in all usage snippets.

1.18.2 — 2026-07-23

• Reframes the watch-scope findings as a review prompt: leaving human-only files (issue templates, changelogs) unwatched is named as a valid choice instead of an implied defect.
• Documents the gate's honest boundaries in the README, on the site, and in the CI gate reference: it verifies link wiring, never whether the root context file actually instructs routing (that is what local telemetry proves), and a root entry-file existence check is named as a planned addition.

1.18.1 — 2026-07-23

• Names the markdown files outside the watch scope as a fourth offender category, so a lowered watch-scope component is as actionable as the other three.
• Writes the gate verdict — score table, findings, pass/fail — to the GitHub run's step summary when CI provides the hook, so reviewers see why a score is not 100% without opening the job log.

1.18.0 — 2026-07-23

• Adds tt gate / /tt gate: a deterministic, telemetry-free CI gate for static documentation discoverability — router coverage, orphaned docs, folder entry points, and watch scope — with a committed no-regression baseline, an optional absolute threshold, shields.io badge output, and exact per-file findings. Discoverable never means discovered: read telemetry stays local and is not used.
• Ships a composite GitHub Action at the repository root (uses: Hedde/trigger_tree@v1.18.0) that installs its own pinned version inside the runner and uploads nothing; any other CI uses pip install trigger-tree.
• Eats the dogfood: this repository runs the gate through the published action on every build, commits its 95% baseline, links the /tt command contract from the docs router (fixing its own orphan finding), and publishes a live docs-discoverability badge alongside coverage.
• Showcases the gate on the website with a self-contained baseline pill that a test locks to the committed .trigger-tree/gate.json, so the shown score can never drift from the enforced one.

1.17.0 — 2026-07-22

• Publishes an example insights report on the website, generated from fixed synthetic telemetry by a committed make demo-report generator and clearly labeled as demo data on the page itself; real reports remain local-only.
• Stops the /tt suggestions skill from repeating output the tool result already shows: the script gains --no-apply-prompt, the Claude contract now answers with a single apply-question line, and protected-summary reasons aggregate reference counts as "heavily referenced" instead of listing every count.

1.16.0 — 2026-07-22

• Restructures /tt suggestions into tiers: at most five numbered, appliable router edits (verified add-link gaps, unlisted router members, missing folder entry points, each with evidence numbers), up to two unnumbered "Worth a look" observations for telemetry signals that need judgment, and one summary line for low-read but likely-critical files — the apply-prompt now appears only when at least one numbered edit exists.
• Classifies .claude/rules/** as always-loaded context: Claude Code injects rules into the system prompt, so Read telemetry can never observe them and they no longer appear as untouched review candidates.
• Skips template files in unlisted-member proposals, deduplicates the same path across suggestion categories, guards empty folders out of cold-folder lines, and marks warming-stage telemetry observations as early signals.

1.15.1 — 2026-07-22

• Fixes a Codex dashboard regression where the detached watcher could bind to the installed plugin directory instead of the user's project, showing zero telemetry while events were being recorded correctly: every entry script and the launcher now honor an explicit TT_PROJECT_DIR, and the Codex skill contract runs all commands from the project directory with that override set, never from the plugin root.

1.15.0 — 2026-07-22

• Replaces the README hero with a real terminal recording of the live dashboard (regenerable via make demo-gif and the committed VHS tape) and embeds the same footage as a compact self-hosted video on the website, clearly labeled against the interactive mock demo.
• Adds one-click copy buttons to the website install blocks and documents the uvx --from trigger-tree tt zero-install route beside pipx in the README and on the site; the site remains free of external requests.

1.14.1 — 2026-07-22

• Brings the website demo to parity with the real dashboard controls: the advertised [s] settings panel now opens and switches the demo's prompt-privacy mode for future prompts, maintenance tips rotate on the real 30-second cadence, recently active folders bubble to the top of the focus sort and settle back after eight seconds, and always-loaded context appears with its injected label instead of being omitted.

1.14.0 — 2026-07-22

• Ships trigger-tree as a standalone PyPI package: pipx install trigger-tree provides a tt console command dispatching to the same bundled scripts and hook manifests, for CI use, git-hook ingestion, and dashboards without a plugin.
• Publishes to PyPI automatically on version tags through trusted publishing (OIDC, no stored secrets) after tests and release integrity pass; release integrity now also verifies the packaged version, and the README, site, and development docs document the standalone install.

1.13.0 — 2026-07-22

• Publishes GitHub release notes automatically from the changelog on every version tag: idempotent create-or-update, normalized vX.Y.Z titles, and a fixed install footer after tests and release integrity pass.
• Adds search presence to the static site — a canonical URL, SoftwareApplication structured data, a sitemap and robots file, and three honestly phrased common questions — with zero new external requests.
• Keeps distribution submission drafts and the weekly scout prompt as local, gitignored material; automated community posting remains explicitly out of scope.

1.12.0 — 2026-07-22

• Records that installation may store a recognizable, gitignored 200-character prompt preview before setup, and states that behavior plainly across the README, site, FAQ, privacy, and security docs.
• Recognizes root GEMINI.md alongside CLAUDE.md and AGENTS.md, and classifies root or nested Gemini context as injected/always loaded rather than an untouched candidate; index.md remains an ordinary folder router.
• Adds a CI and tag-time documentation-currency guard for command parity, top changelog/version parity, and relative-link integrity, with induced-drift tests.
• Removes the obsolete hidden setup flag, standardizes visible search terminology, recompresses the report capture, and verifies the recent release-tag integrity chain and tracked-artifact hygiene.

1.11.1 — 2026-07-22

• Makes recognizable, locally stored 200-character prompt previews the consistent plugin and runtime default instead of falling back to hashed prompts before setup.
• Asks new interactive setups to choose truncate, hash, or off, recommends truncate, keeps automation non-blocking, and preserves existing project choices unless a mode is passed explicitly.
• Updates both Claude and Codex setup workflows to ask for the privacy choice and pass it explicitly, with matching configuration and security documentation.

1.11.0 — 2026-07-22

• Turns the local insights output into a self-contained evidence report with a grade-first hierarchy, compact navigation, generation metadata, and the same section order as /tt insights, while retaining every underlying table.
• Adds pure-Python SVG sparklines, separate count and ratio trend charts, neutral note ticks, small-sample dashes, and a maturity-gated indented documentation tree; cold-start, single-bucket, outlier, long-name, and 100+-file cases are covered.
• Validates the shared five-step heat palette under protan, deutan, and tritan simulations in CI, and flattens report and site treatments to avoid decorative rails, rounded evidence cards, and other visual noise.
• Saves an inspected render of the generated report for documentation reuse while deliberately retaining the clearer synthetic dashboard demo instead of claiming mock footage is a real terminal recording.
• Moves docs-health badge publication to an explicit local-evidence command, make badge-publish; CI now updates measured coverage without overwriting that local aggregate with a permanently immature checkout result.

1.10.0 — 2026-07-22

• Reframes the project around its measurable documentation-discovery loop with a short demo-first README, a routed reference tree, canonical glossary, and concise root CLAUDE.md, while preserving every measurement and platform caveat.
• Adds /tt badge and tt-stats.py --badge, which atomically write a shields.io endpoint and publish only measuring… until the dataset is mature; CI now carries both coverage and docs-health endpoints on the badges branch.
• Makes empty dashboards explain how evidence appears, aligns the HTML report with the product heat palette and grade-first hierarchy, and adds mature cached grades to the statusline without recomputing stats on every refresh.
• Keeps the interactive mock TUI as the canonical demo, adds an optimized static dashboard capture and regenerable VHS tape, and gives the site an honest health story, comparison positioning, privacy statement, accessible contrast, and reduced-motion-safe additions.
• Adds a 1200×630 social card plus Open Graph and Twitter metadata, with no external site requests or analytics, and corrects cross-platform badge tests discovered by the full Windows CI matrix.

1.9.1 — 2026-07-22

• Fixes a critical v1.9.0 regression where Claude Code could load both bundled hook files and double-record session, prompt, and test events. Explicit client markers now make the Codex hook file a silent no-op under Claude while preserving stale-file compatibility and tool-event idempotency.
• Makes the test suite hermetic when trigger-tree's own shell-capture variables are present in the developer environment; the hostile-environment run remains at 100% coverage.
• Restores /tt to explicit user invocation and documents that Windows Claude hooks require a python3 command on PATH, because the documented hook schema has no operating-system condition for exec-form handlers.

1.9.0 — 2026-07-22

• Consolidates Claude telemetry behind its declared hook manifest, routes both clients through one adapter, and suppresses duplicate tool events with a bounded per-session idempotency window.
• Uses Claude's documented shell-free hook execution across platforms and separates the Claude /tt contract from the Codex-only trigger-tree skill package.
• Audits the effective documentation watch scope during setup and doctor runs, proposes observed-layout regexes without applying them implicitly, and distinguishes static hook-file integrity from current, recent, stale, or never-seen hook activity.
• Counts conservative static directory prefixes for pathless Glob and Grep file-glob calls, keeps direct Claude and Codex events in one git-root dataset, and broadens MCP capture while retaining explicit local-file and non-HTTP filtering.
• Adds /tt uninstall to remove only trigger-tree's copied statusline and registration; telemetry data and gitignore entries always remain for explicit manual deletion.
• Updates the README, command contracts, packaging smoke checks, and website to describe the verified behavior and native Windows hook-execution boundary accurately.

1.8.3 — 2026-07-21

• Prevents report and session-cache symlink escapes: HTML reports are written through private atomic files, project-controlled symlink parents are refused, and lock/cache paths no longer follow links outside the repository.
• Keeps runtime shell capture correct and cheap: reads after cd resolve from the real working directory, unsupported POSIX shells retain fallback telemetry, and irrelevant reader commands avoid starting Python when the watched extension is known.
• Uses safe launcher files for Terminal.app paths containing quotes, and consistently prefers Codex-native client markers over Claude compatibility aliases. Launcher creation uses one portable template across BSD/macOS and GNU/Linux mktemp.
• Rejects structurally corrupt current-schema events instead of crashing reports, and excludes symlinked documentation and history files from local analysis.
• Bounds co-read pair generation and watcher prompt state, streams watcher startup history, and avoids rescanning every archive for genuinely new sessions.
• Clarifies that hooks store paths and metadata while selected documentation content is read only for local analysis and is never stored in telemetry or uploaded.
• Exercises symlink-refusal regressions through portable metadata simulations on every CI platform, including Windows environments without symlink privileges.

1.8.0 — 2026-07-21

• Reconciles every current-coverage denominator: the inventory is explicitly split into evaluable and always-loaded files, and touched plus untouched always equals the evaluable total.
• Separates retired telemetry paths from current heat, coverage, and health while retaining their historical reads and per-agent attribution; folder summaries add a measured retired-read share and median current-file age.
• Classifies a discovered folder router as a router even when its filename starts with an underscore, reports unread routers directly, and uses one full reference graph for both inbound counts and displayed samples.
• Compresses untouched review into one ranked table with folded detail and one caveat, hides redundant heat windows and empty folder rows, labels small trend samples, and exposes main-versus-subagent reads in the current heat table.
• Filters client-injected envelopes and command-only noise from task-cluster examples, falling back to the latest real user prompt or no example at all.

1.7.1 — 2026-07-21

• Makes insights proposals evidence-thresholded rather than quota-shaped: zero is a valid result, and every proposed link must be re-verified against two existing files and an absent router reference.
• Adds direct folder-router coverage using the repository's actual README.md, _index.md, index.md, or CLAUDE.md; unrelated in-links no longer hide files that are absent from their own folder entry point.
• Reclassifies scans as non-causal search activity with tool mix, session reach, maximum-session share, and concentrated/distributed patterns. Concentrated bulk work no longer creates router suggestions or lowers the documentation health score.
• Updates the HTML report, README, website, and live terminology to explain the tighter evidence boundaries without adding any project-specific context.

1.6.5 — 2026-07-21

• Excludes injected instruction files and injected-only folders from cold sorting; always-loaded context remains visible in focus/name views without being mislabeled as untouched or cold.
• Preserves the explicit Claude/Codex tip identity and cache-path fallback from the unreleased 1.6.4 candidate.

1.6.4 — 2026-07-21

• Fixes empty dashboard tips by passing Claude/Codex identity explicitly from each command contract into the detached watcher process.
• Adds installed-cache-path detection as a fallback plus regression coverage for both Claude and Codex launch paths.

1.6.3 — 2026-07-21

• Keeps the live dashboard footer visible when a busy tree overflows the terminal: the rotating tip, sort controls, and navigation help can no longer be clipped by the generated files hidden and focused-folder summary rows.
• Adds a crowded-dashboard regression fixture covering the exact failure mode.

1.6.2 — 2026-07-21

• Moves maintenance tips into the live dashboard as a quiet, deterministic line that rotates every 30 seconds; Claude and Codex receive only their own relevant guidance.
• Keeps /tt tips available for direct use but removes it from the prominent command overview. The statusline remains stable and dedicated to current telemetry.
• Passes client identity through every terminal launcher and adds regression coverage for detection, rotation, prompt-history suppression, and launcher quoting.

1.6.1 — 2026-07-21

• Fixes every Claude /tt subcommand resolving scripts relative to the nested command skill instead of the plugin root. /tt watch now opens directly without a failed command and filesystem search first.
• Extends source and isolated-install regression checks so every documented command references a real packaged script through CLAUDE_PLUGIN_ROOT.

1.6.0 — 2026-07-21

• Distinguishes injected instruction context from thermal activity: CLAUDE.md, AGENTS.md, and configured always-loaded paths are labeled injected and cannot misleadingly appear as cold documentation.
• Adds in-dashboard prompt privacy settings (s) with atomic, symlink-safe updates; makes n genuinely toggle A–Z/Z–A and aligns root/nested heat columns exactly.
• Makes the live watcher bounded by file count, preserves partial JSONL appends, and reduces inventory walks without sacrificing deletion detection.
• Keeps statusline totals across history rotation through private per-session summaries, serializes concurrent hook writes, and avoids rescanning the growing history on every refresh.
• Hardens telemetry against symlink writes and terminal escape injection, makes setup writes atomic, applies private permissions, and preserves Bash/zsh reader aliases.
• Adds Python fallback without retrying failed hooks and expands regression coverage for every behavior above across the supported platforms.

1.5.0 — 2026-07-21

• Makes prompt history recognizable after setup: /tt setup now creates a project config using local, gitignored 200-character previews. /tt setup hash, truncate, and off explicitly switch modes; existing history is never rewritten.
• Preserves the privacy-safe installation boundary: before setup, or without a readable project config, prompt logging still falls back to hashes and stores no prompt text.
• Reworks dashboard temperature into a coherent blue → cyan → green → amber → red spectrum. Untouched stays neutral gray, folder colors reflect aggregate heat, and a permanent adaptive heat legend separates temperature from lifetime read counts.
• Updates the README, privacy policy, Claude/Codex workflows, and website demo while preserving the website's six-card, three-column desktop USP grid.

1.4.0 — 2026-07-21

• Adds concise, read-only maintenance tips with strict client separation: Claude Code receives /memory, CLAUDE.md, and scoped-rule guidance; Codex receives AGENTS.md and reproducible-environment guidance. Neither client sees advice for the other.
• Grounds tips in repository evidence such as missing or oversized instruction files, unscoped Claude rules, and absent Codex verification commands. Output is deterministic, bounded to four tips, and never edits instruction or memory files automatically.
• Documents the official Anthropic and OpenAI guidance behind the recommendations and adds /tt tips to the Claude command and equivalent Codex natural-language workflow.

1.3.2 — 2026-07-20

• Restores Claude Code startup and /tt discovery after the withdrawn v1.3.1 package: shared hooks now use the Claude-compatible plugin-root variable that Codex also provides, and the Claude command has an explicit conventional skills/tt/SKILL.md.
• Prevents duplicate shared Claude events while retaining Claude-only skill, failure, and session-end telemetry. Codex-only Stop outcomes no longer run in Claude.
• Adds installed-package regression checks for the command skill and hook environment.

1.3.1 — 2026-07-20

• Clearly separates Claude Code and Codex installation, invocation, hooks, and statusline behavior in the README and website.
• Explains that direct GitHub marketplace installation is immediately usable but is distinct from an OpenAI Curated listing, which requires submission and review.
• Keeps the website's six USP cards in the fixed three-column desktop grid.

1.3.0 — 2026-07-20

• Adds a native Codex plugin manifest, marketplace catalog, trigger-tree skill, and lifecycle hooks while retaining the existing Claude Code plugin in the same repo.
• Normalizes Codex unified-exec cmd payloads, native reads, and filesystem MCP reads into the shared telemetry schema. Repository-root resolution keeps one dataset when Codex starts in a nested directory; hook failures remain silent and non-blocking.
• Documents and tests both installation paths. The website remains exactly six USP cards in a three-column desktop grid.

1.2.2 — 2026-07-20

• Keeps the v1.2.1 dedicated sort-legend fix while replacing the replay integration test's wall-clock race with a deterministic clock. Slow or suspended macOS runners can no longer miss the replay branch and report platform-specific 99.8% coverage.
• v1.2.1 was withdrawn because its tag CI was not fully green. v1.2.2 is the first valid release of the permanent wide/compact dashboard sort legend.

1.2.1 — 2026-07-20

• Moves the live sort controls onto a permanent dedicated legend row instead of sharing space with the heartbeat. Wide panes show [f] focus, [h] hot, [c] cold, and [n] A–Z; narrow panes get an equally complete compact legend.
• Keeps the active sort:<mode> visible and moves prompt navigation, quit, and heartbeat state to their own row, so no sort key disappears in normal split panes.

1.2.0 — 2026-07-20

• Makes historical prompts identifiable without weakening the privacy default: hashed prompts show their stable short ID, marker-only prompts say text is off, and opt-in truncated prompt previews expand with the terminal up to 120 characters.
• Replaces vertical spark glyphs with comparable horizontal five-cell heat bars and moves the heat/lifetime column to the available right edge, making better use of wide split panes while retaining safe narrow-pane truncation.
• Adds explicit live sorting: f recent-focus, h hottest, c coldest (including untouched inventory), and n A–Z. The footer always exposes the keys and active mode; prompt browsing remains a separate bounded chronological timeline.
• Updates the README, plugin help, and six-card/three-column interactive website demo to match the new dashboard behavior.

1.1.0 — 2026-07-20

• Separates current documentation attention from lifetime popularity. Every timestamped read now contributes exponentially decaying heat with a 30-day half-life, while lifetime reads and last-read evidence never disappear.
• Adds deterministic 7-, 30-, and 90-day windows plus file and folder heat to the stats contract and HTML report. Future clock skew is bounded; undated legacy reads remain in lifetime totals without being presented as current heat.
• Updates the live watcher to color and sort by current heat while showing heat and lifetime counts side by side. Prompt browsing remains an exact per-prompt view.
• Makes the watcher a true full-screen TUI: it clears inherited scrollback, disables autowrap while active, and restores the terminal on every exit, preventing refresh frames from accumulating above the dashboard.
• Updates the plugin instructions, README, and six-card/three-column website example to explain the temporal model and its safety boundary: cold means inactive now, never unimportant or safe to remove.

1.0.4 — 2026-07-19

• Fixes an urgent zsh compatibility regression in the runtime reader wrapper: status is a read-only zsh parameter, so assigning it after a successful reader command skipped telemetry and changed the command result to exit 1. The wrapper now uses a private, shell-neutral exit-code variable.
• Runs the exact variable/command-substitution/loop capture regression in both bash and zsh, proving that reader output and exit status remain unchanged while expanded documentation paths are recorded.

1.0.3 — 2026-07-19

• Captures the expanded runtime arguments of successful Bash cat, head, tail, non-mutating sed, and awk calls. Documentation reads through variables, command substitutions, loops, and shell globs now produce exact file-level read events without storing commands, output, patterns, or contents.
• Installs the capture functions through Claude Code's session-scoped environment preamble without rewriting commands or changing permissions. Environments without Bash retain the conservative literal-path parser, and PostToolUse suppresses duplicate reads when runtime capture is active.
• Updates the README, privacy contract, and website copy. The website remains six USP cards in a fixed three-column desktop grid with responsive two/one-column fallbacks.

1.0.2 — 2026-07-19

• Keeps the statusline live for scan-only discovery: scan events now participate in freshness and latest-path display, render as folder paths with a trailing slash, and get their own counter while file/folder/depth totals remain read-derived.
• Records existing watched documentation files consumed by Bash cat, head, tail, non-mutating sed, and awk commands as ordinary file-level reads. Multiple file arguments are preserved; sed -i variants and non-doc files are ignored. End-to-end coverage proves these reads flow through the live watcher.

1.0.1 — 2026-07-19

• Makes the advertised critical-tag and widely-linked safety protections explicit, isolated regression contracts. Critical tags are verified without another protection as a fallback; widely-linked docs are verified at the three-in-link threshold. Both must remain protected, rare-but-critical review items.

1.0.0 — 2026-07-19

• Promotes the verified 1.0 release-candidate contract to stable: context-aware always-loaded classification, safe rare-critical review candidates, history schema migration, hardened diagnostics, and the opt-in experimental outcome view.
• The stable tag is held to the same Linux/macOS/Windows, Python 3.10–3.13, 100%-coverage, formatting, shell, workflow-security, plugin-install, and release-integrity gates used for the release candidate.

1.0.0-rc.1 — 2026-07-19

• Recursively resolves the CLAUDE.md @import graph and classifies injected files as always-loaded instead of cold. Subagent reads retain agent_id/agent_type, while stable tool-call IDs prevent replayed events around compaction from inflating counts.
• Reframes untouched documentation as review candidates. Always-loaded files, safety paths, high-in-link documents, configured critical globs, critical tags, and templates are protected with an explicit likely-keep explanation and a rare-but-critical caveat.
• Adds history schema version 1. Legacy schema-less events migrate explicitly in memory; unknown future versions are rejected and surfaced rather than misread.
• Expands /tt doctor with exact hook-route, config, supported-Python, rotated-log, corrupt-line, legacy-migration, and future-schema diagnostics. Rotation now keeps every archive even when multiple rotations happen in the same second.
• Adds an opt-in TT_EXPERIMENTAL_OUTCOMES='on' view joining documentation reads with local commit and test-command signals. It is labeled correlational—not causal— and remains disabled by default.

0.8.0 — 2026-07-18

• Trustworthy compatibility contract: supported end-user runtimes are now the security-supported Python 3.10–3.13 range, verified independently in CI alongside the Ubuntu, macOS, and Windows test matrix and a strict 100% coverage gate.
• Added Black and Ruff gates, pinned development/test dependencies, and documented a venv-only contributor workflow. Runtime hooks remain dependency-free stdlib Python.
• Added actionlint and zizmor workflow audits, immutable action pins, least-privilege job permissions, and Dependabot coverage for Actions, Python, and npm dependencies.
• Claude plugin validation now uses a locked CLI on Node 22, rejects incomplete npm installs, and proves a real marketplace install inside an isolated Claude config.
• Release tags now fail CI unless the tag, plugin manifest, marketplace, and changelog agree. README and website trust claims now describe the gates that actually run.

0.7.5 — 2026-07-18

• Privacy-hardened prompt markers: the default is now hash, so prompt text is never stored unless a project explicitly opts in to truncate. off remains available for marker-only telemetry; existing project overrides keep their choice.
• Clarified that trigger-tree records local telemetry but sends nothing off-device, and aligned plugin, marketplace, README, security, privacy, and website copy around “See which docs your AI actually discovers.”

0.7.4 — 2026-07-18

• The live overview is now focused on at most ten folders with proven activity: recent activity first, then cumulative reads/searches. Untouched folders and files collapse into one quiet/unread summary instead of filling the viewport.
• Prompt-history remains exact and /tt insights retains the complete cold-path inventory, so focus does not discard evidence.

0.7.3 — 2026-07-18

• The live dashboard now brings the currently active and recently touched folders to the top of the viewport. After eight seconds they return to deterministic alphabetical order; files inside each folder never jump around.
• Prompt-history views remain strictly alphabetical for predictable comparison.

0.7.2 — 2026-07-18

• Updated the CI toolchain to the current Node 24-based GitHub Actions majors: actions/checkout@v7, actions/setup-python@v6, and actions/setup-node@v7. This removes the Node 20 deprecation warnings from every CI job.

0.7.1 — 2026-07-18

• Deleted docs no longer haunt the live tree. Historical Read events retain their counters but cannot re-add a path that no longer exists; historical scans of removed folders likewise stay out of the live overview.
• The watcher refreshes its documentation inventory every second, so a file or folder removed while the pane is open disappears without a restart. Prompt history still shows the original evidence when browsing ←/→.

0.7.0 — 2026-07-18

• Folders no longer look untouched while being searched. Every folder row now keeps 🔍 N searches separate from N unread: scans increase the first counter; only actual Read events lower the second. Scan-only prompts get a visible folder row, and ←/→ browsing recalculates both signals for the selected prompt.
• Concise /tt suggestions. A deterministic formatter captures full stats internally and writes only its evidence scope plus at most five prioritized router edits to stdout. It explicitly says nothing changes until confirmation; cold-start and no-finding states stay one or two lines.
• The interactive website demo mirrors the new folder counters.

0.6.2 — 2026-07-18

• End-to-end proof for Bash discovery telemetry. A cross-process regression test now proves the complete live progression: the watcher starts at 0 reads · 0 scans, an external Bash hook moves it to 0 reads · 1 scan, and only a later Read hook moves it to 1 read · 1 scan.
• The hook manifest itself is regression-tested, ensuring Bash, native search/read, and Skill events cannot silently lose their logger route.
• Arrow escape decoding now tolerates briefly delayed bytes on loaded terminals; the previous 30 ms window proved flaky on a macOS CI runner and could recreate the same ignored/right-arrow symptom under local system load.
• Cross-platform CI now runs every matrix leg even if another platform fails, so release evidence is complete rather than hidden by fail-fast cancellation.

0.6.1 — 2026-07-18

• Bash doc searches are visible. PostToolUse telemetry now recognizes executed rg, grep, and find command segments when they explicitly target existing documentation paths. They appear as scans, never reads, so the live prompt view no longer reports 0 scans while the transcript visibly contains rg lookups.
• Classification is deliberately conservative and private: commands, patterns, output, nonexistent paths, source-code searches, and quoted mentions of rg are not logged. Repeated search segments targeting the same folder in one Bash call produce one scan.

0.6.0 — 2026-07-17

• Prompt navigation is directional and predictable. ← always moves older and → always moves newer; both ends clamp instead of wrapping or changing mode, and a explicitly returns to the live overview. The input decoder now also handles fragmented, application-cursor, and modified arrow escape sequences atomically.
• /tt doctor removes setup guesswork. One command verifies the hook manifest, local-data gitignore, optional statusline, and valid telemetry in the invoking repository, with a concrete recovery action for each problem.
• Real-time means the invoking repo, without path exceptions. The split launcher safely handles spaces, apostrophes, and shell metacharacters. Cross-process tests prove a running watcher displays hook events appended after startup and remains bound to the exact repository.
• 100% is now the release floor. CI raises its cross-platform coverage gate from 80% to 100%.

0.5.3 — 2026-07-17

• Fix: both arrow keys browsed backwards. The key loop mixed select() on the raw fd with buffered sys.stdin.read(): the buffered read slurped the whole escape sequence, the follow-up select() saw an empty fd, and the leftover [ replayed on the next keypress — so ← and → both acted as "prev", one press late. Keys are now read raw via os.read() (read_key()), with a pipe-based regression test.

0.5.2 — 2026-07-17

• Arrow keys now browse prompts too (←/→ on macOS/Linux escape sequences and Windows console codes), and the hint lines say so in plain words ("←/→ browse per prompt"). Fixes a truncation assertion in the 0.5.1 test run.
• Bounded prompt detail: the browser keeps the last 20 prompts (500 events each) while dashboard totals keep aggregating the full history — endless sessions can't grow memory.
• Website demo now demonstrates per-prompt browsing with ←/→ as well.

0.5.1 — 2026-07-17

• Typed prompts are now instantly visible in the live view: every prompt appears in the ticker (▸ "…") the moment you hit enter, and the footer counts prompts alongside reads — so "5 prompts · 0 reads" self-explains an injected-context session instead of looking stale.

0.5.0 — 2026-07-17

• Per-prompt browsing in the dashboard: [ and ] step through every typed prompt; the tree filters to exactly what that input aggregated (reads, scans, skill uses) with the prompt text in the header; a returns to live. Reads before the first prompt land in a "(session start)" bucket.
• End-to-end regression guard: a test now spawns the watcher and the logger as separate processes (exactly how real hooks write) and asserts the live tail picks the events up — the "is it updating?" question is CI-proven forever.

0.4.2 — 2026-07-17

• Liveness heartbeat in the dashboard hint line: "live · last event 12s ago", or "listening for doc reads (injected context never shows here)" when nothing has arrived yet — answers "is it stale or is nothing happening?" at a glance.

0.4.1 — 2026-07-17

• Launcher pane messages now explain exit causes: SIGTERM/SIGINT (e.g. an external kill) reads "terminated from outside — not a crash" instead of a bare status code; real crashes say "crashed". Ends the status-143 confusion.

0.4.0 — 2026-07-17

• External ingestion adapter: tt-log.py ingest '<event-json>' — a stable, validated entry point so any tool (a Codex wrapper, a git hook, an editor plugin) can append telemetry to the same history. Missing ts/session stamped, invalid events dropped silently. Completes the plugin side of multi-tool support.

0.3.7 — 2026-07-17

• /tt watch confirmation line now prints the running plugin version ("trigger-tree v0.3.7 watcher opened …") so a stale session cache is instantly visible. FAQ entry added for the flash-and-disappear symptom (stale pre-v0.3.3 cache → /reload-plugins).

0.3.6 — 2026-07-17

• Positioning, research-backed: README and site now state the validated gap — agent observability tools measure tokens/traces, none measure per-task doc reads; Anthropic's guidance says to prune CLAUDE.md but nothing validates the pruning; the Fallow analogy (unused code → unused docs). Added a who-is-it-for table and an honest Limitations section (official-plugin convention).

0.3.5 — 2026-07-17

• Branding: lowercase everywhere. The plugin now presents itself as trigger-tree across README, website, dashboard banner, SKILL, report and manifests — matching the naming convention of official Claude Code plugins (commit-commands, security-guidance, …). The /tt command is unchanged.

0.3.4 — 2026-07-17

• Stability: layered, crash-proof config resolution in all scripts (project override → plugin default → hardcoded; broken regexes skipped). Fixes the instant-crash of /tt watch in projects with a partial .trigger-tree/config.sh.
• Multi-session safe splits: the iTerm2 split now targets the exact session that invoked it (via ITERM_SESSION_ID) — two Claude sessions in different projects each get their own split, never the frontmost window by accident.
• Discovery-structure awareness: folders without an entry point are flagged ("no index file") and /tt suggestions proposes adding one; _template.md-style files are auto-classified as intentional archive; nested CLAUDE.md/ CLAUDE.local.md files count as always-loaded (they inject on demand and are invisible to read-telemetry — per official memory docs).
• README/site: layman-first hero, "Why measure documentation?" and a sourced "Structuring your docs for discovery" guide; site now has six feature cards and a "Why this matters" section.

0.3.3 — 2026-07-17

• Fix: iTerm2 split closed instantly. iTerm2's AppleScript command parameter is exec-style (no shell), so the compound cd … && … command died on launch. The split now runs a generated launcher script (shebang + shell), which also keeps the pane open with the error message if the watcher ever exits non-zero.

0.3.2 — 2026-07-17

• Dashboard visual parity with the website demo: three-tier heat palette (green → amber → red), clean folder lines (untouched counters only appear when files are collapsed), "just now" ticker, dim q quit hint line.
• Documentation health grade: deterministic A–F score (coverage, router gaps, hunting ratio) in stats, /tt status, /tt insights and as hero card in the HTML report — one number a product owner can track sprint over sprint.

0.3.1 — 2026-07-17

• /tt watch on macOS now stays in the terminal you called it from: iTerm2 users get a split pane in the current window (detected via TERM_PROGRAM), with a new-window fallback; Terminal.app remains the fallback for everything else.

0.3.0 — 2026-07-17

• Windows support: posix path normalization, utf-8 everywhere, ANSI console enable + msvcrt keys in the dashboard, wt.exe/start launcher branch, python3||python fallbacks. CI test matrix now runs ubuntu + macos + windows.
• Folder heat & cold map: per-folder coverage and read volume in stats and the HTML report.
• Router-gap detection: untouched files are cross-referenced against all docs — "untouched and unreferenced" pinpoints where the router is blind.
• /tt suggestions: max 5 prioritized, evidence-backed router fixes, applied only after confirmation.
• README overhaul (hero, how-it-works with hook transparency, FAQ, platform matrix) and a website at hedde.github.io/trigger_tree — an interactive recreation of the live dashboard, with a few easter eggs for the curious.

0.2.1 — 2026-07-17

• Rename plugin id tt → trigger-tree for the plugin directory (unique, descriptive name). The /tt command is unchanged — it comes from the root skill's name field.
• GitHub Pages docs site: https://hedde.github.io/trigger_tree/

0.2.0 — 2026-07-17

• Skill-tool telemetry: PostToolUse hook on Skill logs skill invocations; an invoked skill's SKILL.md counts as touched, shrinking the always_loaded blind spot.
• Portability: logger and statusline rewritten in pure python3 (jq and BSD date dependencies dropped); /tt watch opens a tmux split, macOS Terminal, or gnome-terminal/konsole/xterm.
• Prompt privacy: TT_LOG_PROMPTS=truncate|hash|off — fingerprints keep working in all modes.
• Log rotation: history.jsonl rotates to timestamped archives beyond TT_ROTATE_BYTES (default 5 MB); aggregator and watcher read all archives.
• Trend + annotations: daily/weekly reads-scans buckets with hunting ratio; /tt note <text> marks router changes on the timeline so their effect is visible.
• Task clusters: fingerprints grouped by Jaccard similarity (≥ 0.6) instead of exact-set matching.
• /tt setup: idempotent project wiring (gitignore, statusline copy + registration, optional config override).
• CI: unit + smoke tests with an 80% coverage gate, shellcheck, claude plugin validate. Measured coverage is published as a live badge (shields endpoint JSON on the badges branch — no external coverage service).
• Community files: CONTRIBUTING, code of conduct, security policy, issue/PR templates, README badges.

0.1.0 — 2026-07-17

Initial release.

• Telemetry hooks (SessionStart, UserPromptSubmit, PostToolUse on Read|Glob|Grep) → .trigger-tree/history.jsonl, shell-side, zero model tokens.
• /tt root skill with subcommands: status, watch [demo|replay], insights, help.
• Live ASCII pulse dashboard (tt-watch.py): heat-colored doc tree, white flash + upward ripple per read, hunting ticker, subagent attribution.
• Deterministic aggregator (tt-stats.py): per-file read counts, task fingerprints, co-read pairs, hunting, untouched paths with a maturity model (cold-start → warming → mature).
• Self-contained HTML report (tt-report.py), published via Artifact from /tt insights.
• Statusline script with age-based pulse dot (ships with the plugin; registered per project).