@jjanczur/tyran 0.1.0 → 0.1.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -11,7 +11,7 @@
11
11
  {
12
12
  "name": "tyran",
13
13
  "description": "A task conductor for Claude Code: multi-agent orchestration with an enforced evidence contract, repo-specific learning, and update-safe local evolution.",
14
- "version": "0.1.0",
14
+ "version": "0.1.4",
15
15
  "source": "./",
16
16
  "author": {
17
17
  "name": "Jacek Janczura",
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "tyran",
3
- "version": "0.1.0",
3
+ "version": "0.1.4",
4
4
  "description": "A task conductor for Claude Code: multi-agent orchestration with an enforced evidence contract, repo-specific learning, and update-safe local evolution.",
5
5
  "author": {
6
6
  "name": "Jacek Janczura",
package/CHANGELOG.md CHANGED
@@ -1,6 +1,301 @@
1
1
  # Changelog
2
2
 
3
- ## 0.1.0 — unreleased
3
+ ## 0.1.4 — 2026-08-01
4
+
5
+ ### The gate stopped refusing Claude Code's own memory
6
+
7
+ Adopting Tyran in a repo made the policy gate refuse the harness's OWN writes:
8
+ `Write` to `~/.claude/projects/<slug>/memory/*.md` came back "outside this
9
+ repository, class KERNEL". The memory store and the per-session scratchpad are
10
+ genuinely outside every repository, so `repoRelative` returned null and the
11
+ write matrix answered KERNEL — the same branch that correctly refuses writing
12
+ into somebody else's repo. The cost was that the assistant could no longer
13
+ persist what it learned while working in a Tyran repo, and a boundary that
14
+ blocks the tool's own bookkeeping is one its user turns off.
15
+
16
+ The gate now exempts exactly two locations, and only for the `main` thread:
17
+ `<config>/projects/<slug>/memory/` and `<tmp>/claude-<id>/`. Everything else
18
+ outside the repo is still KERNEL; the rest of the config dir — including
19
+ `settings.json`, which registers these very hooks — stays untouchable; and a
20
+ subagent is never exempted, because it has no business writing outside its
21
+ worktree. Bash never hit this: an out-of-repo token yields no finding, so
22
+ `echo > scratch` was already allowed — the refusal lived only on the
23
+ file-writing tools, which is the asymmetry a user actually meets.
24
+
25
+ ### npm publishes on a pushed tag
26
+
27
+ `npm-publish.yml` now also triggers on a pushed `v*` tag, not only on a
28
+ published GitHub release. npm served 0.1.0 while the marketplace was three
29
+ releases ahead, because the manual release step was skipped for 0.1.1, 0.1.2
30
+ and 0.1.3; a release step a human has to remember is one that drifts. The
31
+ version-must-match-the-tag guard runs for both triggers, and npm refuses a
32
+ duplicate version, so the two are idempotent.
33
+
34
+ ## 0.1.3 — 2026-07-29
35
+
36
+ From two independent field reports on complete initiatives — one M, one L —
37
+ run on 0.1.1 and 0.1.2. Five of the reported items are fixed; the rest are
38
+ recorded here as open, with the reason.
39
+
40
+ ### A git worktree is the same repository
41
+
42
+ Two reports described **opposite** failures with one cause: the policy gate
43
+ read "the repository" as "this directory tree", and a worktree is neither
44
+ inside it nor a different repository.
45
+
46
+ - **Silent.** A session running *in* a worktree found no `.tyran/` there — the
47
+ directory is committed data and `git worktree add` gives a fresh checkout —
48
+ so the gate concluded Tyran had not been adopted and said nothing. Four
49
+ worktrees, four implementers with no autonomy class and no path classes, and
50
+ `git push origin main` passing at every deployment class.
51
+ - **Loud.** A session in the main checkout writing *into* a worktree got
52
+ `normalizePath → null → KERNEL`: every `Edit` refused as "outside this
53
+ repository", while `skills/run/SKILL.md` rule 7 *requires* a worktree per
54
+ agent. Five agents hit it in one initiative and all rerouted through `Bash`
55
+ heredocs — a channel the gate does not class at all. A refusal that moves
56
+ work somewhere less visible is worse than no refusal, and this one was the
57
+ plugin contradicting itself.
58
+
59
+ The gate now asks which repository a path belongs to. Detection is pure
60
+ filesystem — a linked worktree's `.git` is a *file* holding `gitdir:` — so the
61
+ write path still runs no child processes. Policy and config are inherited from
62
+ the main checkout, and the test is **identity, not proximity**: both sides must
63
+ resolve to the same main checkout, so a path in a different repository is still
64
+ KERNEL. Three mutants were installed and killed, including one that survived
65
+ the first version of the push test — it refused for the wrong reason and the
66
+ assertion could not tell.
67
+
68
+ ### `git stash push` was read as `git push`
69
+
70
+ `words.has('push')` was true when the word appeared anywhere in the segment, so
71
+ `git stash push --staged -m "conductor: ..."` — purely local — was classified as
72
+ a publish. The remote was then read as the token after the word: `conductor` out
73
+ of the message, or `2` out of a `2>&1`. The refusal told the operator to run
74
+ `git remote set-head conductor -a`, which cannot succeed, about a repository
75
+ whose default branch was recorded all along.
76
+
77
+ The cost was not one bad refusal: rule 7 *requires* addressed stashes to protect
78
+ other windows' work, so the gate refused the safe half of a workflow the plugin
79
+ mandates and pushed agents toward the deprecated positional form. `git remote
80
+ add` and `git worktree add` had the same shape against the `add` rule — and
81
+ `worktree add` is what rule 7 tells every parallel agent to run.
82
+
83
+ Classification now resolves the **subcommand slot**. It is not a pure narrowing:
84
+ `git subtree push` publishes for real and its verb is the second word, so
85
+ namespace subcommands forward their verb. Dropping that would have traded a
86
+ false positive for a false negative on a genuine push.
87
+
88
+ ### `journal.mjs append` issues the id it used to demand
89
+
90
+ `skills/run/SKILL.md` says IDs never come from memory, because after a
91
+ compaction memory hands out the same number twice — and `append` then *rejected*
92
+ a missing `data.id`, leaving the conductor to remember a separate `next-id`
93
+ call. Measured: 12 decision IDs hand-assigned from memory in one initiative,
94
+ the exact failure the rule names, with nothing objecting. Both reports raised
95
+ it independently. An explicit id still wins.
96
+
97
+ Two errors that could only be answered by grepping the plugin's source now
98
+ answer themselves: a rejected `ev` lists the closed set, and a missing `data`
99
+ key names the whole contract for that event rather than the first key checked.
100
+ `--help` prints the per-event table.
101
+
102
+ ### Worktrees, in the conductor's own words
103
+
104
+ Rule 7 said "a git worktree per agent" and stopped. A fresh worktree has no
105
+ dependencies, so every validation command exits 127 and the evidence contract
106
+ cannot be satisfied by construction. The rule now says to link the main
107
+ checkout's dependency directory — safe precisely because the manifest and
108
+ lockfile are already a shared zone nobody may edit — and notes that
109
+ `.tyran/state/**` exists only in the main checkout.
110
+
111
+ Also: the conductor is told it is notified when a subagent finishes and should
112
+ not poll. On a real run most polled checkpoints fired after the agent had
113
+ already reported, and the operator watched the conductor disown dozens of stale
114
+ notifications.
115
+
116
+ ### Open, and why
117
+
118
+ - **No lease on the checkout.** Two agents in one checkout is the failure rule 7
119
+ exists to prevent, and nothing enforces it; leases cover the heavy slot and
120
+ not the working tree. The reporting retro reached the same conclusion and
121
+ refused to patch it locally, correctly — the only real fix is a hook, and
122
+ `hooks/**` is KERNEL, so no session can build one. It needs a design pass,
123
+ not a hurried one.
124
+ - **The gate matches command TEXT**, so `ls -1 .npmrc` is refused as a
125
+ credential read though it reads nothing, and the scratch directory a `GATED`
126
+ refusal tells an agent to write its diff into is refused as outside the repo.
127
+ The second is self-defeating and should be fixed; both are refusal semantics.
128
+ - **`doctor` does not enumerate worktrees or cross-check `RELEASED` leases**
129
+ against resources that still exist. A field run lost a validation baseline to
130
+ 32,243 phantom lint errors from leftover worktrees.
131
+ - **Forge detection.** Tyran assumes GitHub; on Bitbucket the default terminal
132
+ step of the default autonomy class has no defined path, and two agents each
133
+ rediscovered that.
134
+
135
+ ## 0.1.2 — 2026-07-29
136
+
137
+ From a field report on a real 0.1.1 install. Every item below is a failure
138
+ somebody hit, not one anybody predicted.
139
+
140
+ ### The worktree with no boundary
141
+
142
+ **`git worktree add` carries tracked files only, and the policy gate is
143
+ deliberately silent in a repo with no `.tyran/`.** Put those together and an
144
+ uncommitted `.tyran/` means every worktree the conductor creates has no config
145
+ and no policy — so the agents there run with **no autonomy class and no path
146
+ classes at all**. Nothing is refused, nothing fails, and `git push origin main`
147
+ returns `PASS` on line 1213 of the gate. Measured: four worktrees, four ungated
148
+ implementers, and no way for the operator to see it.
149
+
150
+ This is the worst kind of defect this project can have, because Tyran's whole
151
+ argument is that the boundary is a mechanism rather than a promise — and here
152
+ it was absent precisely where the most agents run.
153
+
154
+ Three changes, none of which touch the gate:
155
+
156
+ - `doctor --state` gains **`tyran-dir-untracked`** (warning): git is asked
157
+ whether anything under `.tyran/` is tracked. It returns *no answer* outside a
158
+ work tree rather than inventing one, the same discipline the lockfile rule
159
+ below uses;
160
+ - `/tyran:setup` now ends by asking for the commit, **with the reason**, rather
161
+ than leaving `.tyran/` sitting untracked;
162
+ - the conductor treats a worktree without `.tyran/` as a precondition failure,
163
+ and is told to fix it by committing on the main checkout — not by copying the
164
+ file in, which just makes four divergent configs one layer down.
165
+
166
+ ### Two inferences that produced a broken gate
167
+
168
+ **A validation command that watches does not fail an agent, it hangs it.**
169
+ Setup wrote `pnpm test` into a repo whose `test` script is bare `vitest` —
170
+ watch mode, no output, no timeout, a session that simply stops. `detectValidation`
171
+ now reroutes to a run-once variant (`test:run`, `test:ci`, `test:once`) and, if
172
+ there is none, **leaves the command out** and flags the config rather than
173
+ writing a booby trap. The rule is deliberately narrow — `jest` runs once by
174
+ default and is not treated as a watcher — because a false positive here silently
175
+ drops a real test command.
176
+
177
+ **A lockfile on disk is not evidence that it is the repo's lockfile.** The same
178
+ install had `pnpm-lock.yaml` gitignored and untracked alongside the tracked
179
+ `package-lock.json` that its deploy actually builds from, and `.gitignore` said
180
+ so in words. Picking by disk order chose pnpm, and then every single validation
181
+ command was wrong. `detectPackageManager` now asks git which lockfiles are
182
+ tracked and says which one it rejected. An empty `git ls-files` means git could
183
+ not answer and changes nothing — "not a git repo" is not evidence that a file is
184
+ ignored.
185
+
186
+ Both replay clean against the repository that produced them.
187
+
188
+ ### Errors that named the wrong cause
189
+
190
+ `yaml-lite` answered every unbalanced quote with *"quote the whole value if it
191
+ contains an apostrophe"*. An operator writing a long `source:` as a multi-line
192
+ double-quoted string was sent hunting an apostrophe that did not exist, three
193
+ times. A value that opens a quote and never closes it now says exactly that,
194
+ and says the subset has no multi-line scalars. The apostrophe advice survives
195
+ only on the case it is actually about.
196
+
197
+ `doctor.mjs` with no flag printed *"--state is the only mode today"* long after
198
+ `--hooks` shipped, sending readers to look for a flag they already had.
199
+
200
+ The generated `config.yaml` header now states the YAML subset and asks for the
201
+ commit, since both traps are sprung by hand-editing a file whose constraints
202
+ were documented somewhere else.
203
+
204
+ ### Still open, deliberately
205
+
206
+ The report also found that the gate matches on command TEXT, so `ls -1 .npmrc`
207
+ is refused as a credential read though it reads nothing, and that a `GATED`
208
+ refusal tells an agent to write a diff for the operator while `outside this
209
+ repository is never autonomous` refuses the scratch directory that is the
210
+ obvious place to put it. That second one is self-defeating and worth fixing.
211
+ Both are changes to refusal semantics in a KERNEL file and are not being made
212
+ in the same pass as everything above.
213
+
214
+ ## 0.1.1 — 2026-07-29
215
+
216
+ Both entries below come from one install of 0.1.0 on a real repository. The
217
+ version is bumped rather than folded into 0.1.0 because `claude plugin update`
218
+ compares **versions**, not commits: with the manifest left at 0.1.0 it reports
219
+ "already at the latest version" and users never receive the fix.
220
+
221
+ ### `.tyran/config.yaml` is AUTO, and what that costs is written down
222
+
223
+ **The file that says how to validate this repo was one an agent could not
224
+ fix.** Measured on the same install as below: setup inferred `pnpm test`,
225
+ which in that repo is bare `vitest` — watch mode, it never exits, and it would
226
+ have hung every agent handed to it. The session that *discovered* that could
227
+ not repair the file that said it. It produced a diff, then a heredoc for the
228
+ operator to paste, then explained why it would not route around its own gate.
229
+ It was right not to. The gate was wrong.
230
+
231
+ So the shipped policy classes `.tyran/config.yaml` `AUTO`. It is mostly a
232
+ description of the repository — package manager, validation commands, shared
233
+ zones — and the agent that finds it wrong is what is best placed to correct it.
234
+
235
+ The cost is real and is now stated in all four places a user might read
236
+ (`templates/policies/autonomy.yaml`, `docs/configuration.md`,
237
+ `docs/policy-gate.md`, `docs/self-improvement.md`), because an unstated cost is
238
+ a false guarantee: **`autonomy:` lives in this file too, so an agent can raise
239
+ its own deployment class from P1 to P3 and then push to main.** What remains is
240
+ weaker and is said as such — nothing *infers* a raise, and a raise is a diff in
241
+ a committed file with its provenance beside it. Note that `docs/policy-gate.md`
242
+ had already measured this exact escalation happening *under* GATED, in an
243
+ unattended main loop, wherever `Write` is allow-listed; GATED was buying less
244
+ here than it looked like it was buying.
245
+
246
+ The push gate's own refusal named the old class too, and naming a class there
247
+ was the underlying mistake rather than naming the wrong one: the text is
248
+ per-repo configuration, so any class it prints goes stale the moment somebody
249
+ reclassifies the file. It now says the policy decides and names nothing, and
250
+ the test asserts the *absence* of a class claim instead of pinning a string
251
+ that has already been wrong once.
252
+
253
+ Two claims that had gone false are corrected rather than left standing:
254
+ `README.md`'s "the config file holding it is GATED, not KERNEL" and
255
+ `docs/configuration.md`'s "Tyran never raises it on its own — that path is
256
+ blocked by the policy hook". A test now pins the class *and* pins that the
257
+ template still spells out what it gives up, because flipping it back is a
258
+ one-word edit nobody would notice in review.
259
+
260
+ ### Setup no longer locks the repository it is setting up
261
+
262
+ **`/tyran:setup` created `.tyran/config.yaml` and stopped, and that one
263
+ missing file made the repository unwritable.** The policy gate is silent in a
264
+ repo with no `.tyran/` directory and refuses every write in a repo that has
265
+ one without `.tyran/policies/autonomy.yaml` under it. Both halves are right;
266
+ their seam was not. Setup's own first command moved a fresh repo from
267
+ "unmanaged" to "refuses everything" — including the write that would have
268
+ installed the missing policy. Measured on a real install: the session ended
269
+ with the operator being handed a `mkdir` and a `cp` to run by hand, in the
270
+ middle of a one-command setup.
271
+
272
+ `scan-repo.mjs` now installs the policy from the shipped template **before**
273
+ it writes the config, and removes what it created if it cannot, so a failed
274
+ bootstrap cannot leave behind the exact state it exists to prevent. A repo
275
+ from before this change is repaired with `--ensure-policy`, which touches
276
+ nothing else.
277
+
278
+ That is bootstrap, not a loop authorizing itself, and the difference is
279
+ mechanical rather than a matter of intent: it only ever creates — an existing
280
+ policy is never read, merged or overwritten — and what it writes is the
281
+ shipped template byte for byte, the strictest default Tyran has. No input
282
+ makes it emit something weaker. Editing the file afterwards is human-only,
283
+ enforced exactly as before. Three tests pin the properties that make that
284
+ sentence true rather than aspirational: byte-identity with the template, an
285
+ untouched hand-written policy across setup and repair, and no `.tyran/` left
286
+ on disk when the install fails.
287
+
288
+ Two smaller things the same incident exposed. `doctor`'s `policy-missing`
289
+ finding was severity `info` — a repository where every tool call is denied,
290
+ reported as a note — and its suggested fix was a `cp` naming the policy path,
291
+ which the gate refuses when an agent runs it; it is now an `error` carrying a
292
+ command that works. And the documented `schema.mjs validate policy
293
+ .tyran/policies/…` line is one a **human** runs: inside a session it names a
294
+ path under `.tyran/policies/**` and is refused like any other. `doctor --state
295
+ --dir .tyran` validates the same file and names only the directory. The docs
296
+ say so now, and `/tyran:setup` runs the latter.
297
+
298
+ ## 0.1.0 — 2026-07-27
4
299
 
5
300
  ### Six protocol skills, and the budget raised once to pay for them
6
301
 
package/README.md CHANGED
@@ -21,11 +21,15 @@
21
21
  <h3 align="center">The more you use it, the better it gets.</h3>
22
22
 
23
23
  <p align="center">
24
- Tyran is a task conductor for Claude Code: it interviews you, plans, and drives a team of
25
- fresh-context agents through your work — <b>refusing to believe any agent that can't show
26
- raw command output as proof</b>, routing every subtask to the cheapest model that can do
27
- it, and keeping each agent in the context band where it still thinks clearly. When the
28
- work is done, it will not let the initiative end until it has learned something from it.
24
+ Never written a skill, spun up a subagent, or run an agent team yourself? Tyran does all
25
+ three automatically, on your behalf — and the retrospective it runs after every initiative
26
+ means it keeps learning your repo's way of working instead of repeating a generic playbook.
27
+ </p>
28
+
29
+ <p align="center">
30
+ Under the hood: it interviews you, plans, and drives a team of fresh-context agents through
31
+ your work, routing every subtask to the cheapest model that can actually do it — so a file
32
+ rename doesn't burn through your subscription at top-tier rates.
29
33
  </p>
30
34
 
31
35
  ---
@@ -142,7 +146,7 @@ Initiative 20 Has written repo-specific skills for your recurring work and
142
146
  ```text
143
147
  /plugin marketplace add jjanczur/tyran
144
148
  /plugin install tyran@tyran
145
- /tyran:setup # scans the repo, writes .tyran/config.yaml, installs /tyran
149
+ /tyran:setup # scans the repo, writes .tyran/, installs /tyran
146
150
  /tyran # then just describe what you want done
147
151
  ```
148
152
 
@@ -158,9 +162,11 @@ Install the Tyran plugin in this repository and set it up for me.
158
162
  1. Run: claude plugin marketplace add jjanczur/tyran
159
163
  2. Run: claude plugin install tyran@tyran
160
164
  3. Tell me to restart Claude Code, so the hooks and agents load.
161
- 4. After the restart, run /tyran:setup. It scans this repo and writes
162
- .tyran/config.yaml. Walk me through what it inferred - especially the
163
- validation commands and the deployment autonomy class - before we commit it.
165
+ 4. After the restart, run /tyran:setup, then /tyran:doctor. Setup writes the
166
+ two files under .tyran/ that Tyran needs - the config, and the autonomy
167
+ policy the write gate reads - and doctor proves the install actually
168
+ works. Walk me through what setup inferred - especially the validation
169
+ commands and the deployment autonomy class - before we commit it.
164
170
 
165
171
  Docs: https://jjanczur.github.io/tyran/getting-started/
166
172
  ```
@@ -278,10 +284,12 @@ echo "wrong branch — hold everything" > .tyran/STOP
278
284
  force-pushes are refused; a push to your production branch is refused at the
279
285
  strictest autonomy class, including through `HEAD`, `@`, a shell variable or
280
286
  a `git -c` alias. **What it does not do is stop an agent from raising the
281
- class itself:** the config file holding it is GATED, not KERNEL, so an agent
282
- with a broad allow-list can edit it in the main loop — measured, not
283
- supposed. Autonomy classes are a convention the gate enforces downward, not
284
- a lock. Not a firewall, and
287
+ class itself:** the config file holding it is `AUTO` in the shipped policy,
288
+ so an agent can edit it. That is deliberate — the same file holds your
289
+ validation commands, and an agent that cannot correct a wrong one is an
290
+ agent that runs your test suite wrong forever. Set the rule back to `GATED`
291
+ in one word if you want the other trade. Autonomy classes are a convention
292
+ the gate enforces downward, not a lock. Not a firewall, and
285
293
  [docs/hooks.md](docs/hooks.md) says exactly where it stops — including the
286
294
  scanner's own measured false-negative rate.
287
295
 
@@ -341,53 +349,48 @@ one directory each.</sub>
341
349
 
342
350
  ## Command-line use (outside Claude Code)
343
351
 
344
- A single `bin/tyran.mjs` entry point exposes the same scripts the plugin's
345
- hooks and skills already call — `doctor`, `scan-repo`, `tiers`, `journal`,
346
- `schema`, `stop-check`, `scan-control-chars`, `desc-budget` — for a shell or a
347
- CI job with no Claude Code session. Zero dependencies; `--help` lists them.
352
+ Published on npm as [`@jjanczur/tyran`](https://www.npmjs.com/package/@jjanczur/tyran)
353
+ — scoped, not the bare name `tyran`, which carries an unpublish tombstone from
354
+ 2021-03-30 that npm's abuse policy blocks anyone from reusing, permanently. The
355
+ command stays `tyran` either way; only the registry name is scoped, the same
356
+ pattern as `@angular/cli` giving you `ng`. `bin/tyran.mjs` exposes the same
357
+ scripts the plugin's hooks and skills already call — `doctor`, `scan-repo`,
358
+ `tiers`, `journal`, `schema`, `stop-check`, `scan-control-chars`,
359
+ `desc-budget` — for a shell or a CI job with no Claude Code session. Zero
360
+ dependencies; `--help` lists them.
348
361
 
349
362
  ```bash
350
- node bin/tyran.mjs doctor --hooks # is any gate installed but unable to fire?
351
- node bin/tyran.mjs scan-repo --dir . # what this repo looks like, with provenance
363
+ npx @jjanczur/tyran doctor --hooks # is any gate installed but unable to fire?
364
+ npx @jjanczur/tyran scan-repo --dir . # what this repo looks like, with provenance
352
365
  ```
353
366
 
354
367
  Exit codes propagate to the digit, so a CI step reddens exactly when the
355
368
  underlying script does.
356
369
 
357
- > **Not on npm yet — so `npx @jjanczur/tyran` does not work today.** The
358
- > package is built, versioned and tested; only the publish is outstanding.
359
- > Saying it works here before that is true would be exactly the kind of claim
360
- > this project refuses to accept from its own agents. It publishes under the
361
- > scope, not the bare name `tyran` — that name carries an npm unpublish
362
- > tombstone from 2021-03-30, and npm's policy against reusing an unpublished
363
- > name is permanent, for anyone, not a grace period this project could wait
364
- > out. The command stays `tyran` either way; only the registry name is scoped.
365
- >
366
- > **And it would not install the Claude Code plugin either.** For that, run
370
+ > **This does not install the Claude Code plugin.** For that, run
367
371
  > `/plugin marketplace add jjanczur/tyran` inside Claude Code — that name is
368
- > the Claude Code marketplace identity, a separate namespace from npm and
369
- > unaffected by any of the above. The npm package is the tooling; the plugin
370
- > is the conductor.
372
+ > the Claude Code marketplace identity, a separate namespace from npm. The npm
373
+ > package is the tooling; the plugin is the conductor.
371
374
 
372
375
  ## Documentation
373
376
 
374
377
  **Everything below also reads as a site: [jjanczur.github.io/tyran](https://jjanczur.github.io/tyran/)**
375
378
  — same text, with search, rendered diagrams and per-page status badges.
376
379
 
377
- - 📖 [Getting started](docs/getting-started.md)
378
- - 🧰 [Skills and agents](docs/skills.md) — all fourteen skills and four agents, what each assumes, and who invokes it
379
- - ⚙️ [Configuration](docs/configuration.md) — `.tyran/config.yaml`, cost profiles, autonomy classes
380
- - 🎭 [The roster and model routing](docs/agents.md) — the four agents, the tier and effort table, dynamic overrides, the `.tyran/STOP` brake (shipped)
381
- - 🏛️ [Architecture](docs/architecture.md) — the three layers, the journal, the hooks
382
- - 📜 [Journal reference](docs/journal.md) — the append-only event schema (shipped)
383
- - 🧾 [Projections](docs/projections.md) — generated `STATE.md` / `PROGRESS.md` and `--check` (shipped)
384
- - 🩺 [Doctor](docs/doctor.md) — `--state` consistency check: drift, orphan leases, dead policy rules (shipped)
385
- - 🪝 [Hook runtime](docs/hooks.md) — gates vs probes, why hooks fail open, and what the deadline really promises (shipped)
386
- - 🧾 [Evidence gate](docs/evidence-gate.md) — the criterion, who it binds, the recorded escape hatch, and the line between silence and forgery (shipped)
387
- - 🛡️ [Policy gate](docs/policy-gate.md) — path classes, the deployment class, the one rule on reads, and where each stops (shipped)
388
- - 🧠 [Self-improvement](docs/self-improvement.md) — how Tyran learns your repo, and its guardrails
389
- - ❓ [FAQ](docs/faq.md)
390
- - 🤝 [Contributing](CONTRIBUTING.md)
380
+ - 📖 [Getting started](https://jjanczur.github.io/tyran/getting-started/)
381
+ - 🧰 [Skills and agents](https://jjanczur.github.io/tyran/skills/) — all fourteen skills and four agents, what each assumes, and who invokes it
382
+ - ⚙️ [Configuration](https://jjanczur.github.io/tyran/configuration/) — `.tyran/config.yaml`, cost profiles, autonomy classes
383
+ - 🎭 [The roster and model routing](https://jjanczur.github.io/tyran/agents/) — the four agents, the tier and effort table, dynamic overrides, the `.tyran/STOP` brake (shipped)
384
+ - 🏛️ [Architecture](https://jjanczur.github.io/tyran/architecture/) — the three layers, the journal, the hooks
385
+ - 📜 [Journal reference](https://jjanczur.github.io/tyran/journal/) — the append-only event schema (shipped)
386
+ - 🧾 [Projections](https://jjanczur.github.io/tyran/projections/) — generated `STATE.md` / `PROGRESS.md` and `--check` (shipped)
387
+ - 🩺 [Doctor](https://jjanczur.github.io/tyran/doctor/) — `--state` consistency check: drift, orphan leases, dead policy rules (shipped)
388
+ - 🪝 [Hook runtime](https://jjanczur.github.io/tyran/hooks/) — gates vs probes, why hooks fail open, and what the deadline really promises (shipped)
389
+ - 🧾 [Evidence gate](https://jjanczur.github.io/tyran/evidence-gate/) — the criterion, who it binds, the recorded escape hatch, and the line between silence and forgery (shipped)
390
+ - 🛡️ [Policy gate](https://jjanczur.github.io/tyran/policy-gate/) — path classes, the deployment class, the one rule on reads, and where each stops (shipped)
391
+ - 🧠 [Self-improvement](https://jjanczur.github.io/tyran/self-improvement/) — how Tyran learns your repo, and its guardrails
392
+ - ❓ [FAQ](https://jjanczur.github.io/tyran/faq/)
393
+ - 🤝 [Contributing](CONTRIBUTING.md) — not a docs-site page, so this one stays a repo link
391
394
 
392
395
  ## Strategic principles
393
396
 
@@ -406,9 +409,9 @@ underlying script does.
406
409
  generalizable improvements travel to the core as pull requests.
407
410
  6. **Autonomy earns trust in layers.** Detected once, confirmed with you, and
408
411
  enforced downward by the policy gate. Not *"never self-escalated"* — the
409
- file holding the class is GATED rather than KERNEL, and an agent with a
410
- broad allow-list can edit it. Measured, not assumed; see
411
- [the policy gate](docs/policy-gate.md).
412
+ file holding the class is `AUTO`, so an agent can edit it, chosen so that
413
+ an agent can also repair the validation commands stored beside it.
414
+ Measured, not assumed; see [the policy gate](docs/policy-gate.md).
412
415
 
413
416
  ## Roadmap
414
417