@awebai/oats 0.45.0 → 0.47.0

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.
@@ -0,0 +1,147 @@
1
+ # Explicit native harness trust
2
+
3
+ Run this command on the host where Claude or Codex runs, while other native
4
+ configuration editors are quiescent:
5
+
6
+ ```sh
7
+ oats harness trust --dir /absolute/deployment --harness all --plan
8
+ oats harness trust --dir /absolute/deployment --harness all
9
+ ```
10
+
11
+ `--plan` reads only. The bare command applies a fresh plan; there is no `--apply`.
12
+ The default harness is `all`, in Claude-then-Codex order. `claude` or `codex`
13
+ selects one. The deployment is resolved from the existing local command context,
14
+ including when invoked from an instance home, then canonicalized with the native
15
+ filesystem lookup. The result, native key, audit and reader verification use that
16
+ on-disk spelling, including on case-insensitive APFS. No remote
17
+ adapter, arbitrary file selector, parent-root substitution or harness launch is
18
+ performed. A native executable need not be installed to edit the qualified format.
19
+
20
+ The command writes only the canonical deployment root's native trust leaf:
21
+
22
+ - Claude: `projects[root].hasTrustDialogAccepted = true` in
23
+ `$CLAUDE_CONFIG_DIR/.claude.json`, or `~/.claude.json`.
24
+ - Codex: `projects[root].trust_level = "trusted"` in
25
+ `$CODEX_HOME/config.toml`, or `~/.codex/config.toml`.
26
+
27
+ Nonempty override directories must be absolute. Directory aliases are resolved;
28
+ symlink/nonregular/hardlinked leaves, foreign ownership and ACLs refuse. On Linux, GNU ls's single `.` security-context marker is accepted;
29
+ the context must be readable, match the same-directory candidate before replacement,
30
+ and remain unchanged at verification. A differing directory-default context refuses
31
+ before rename; OATS does not relabel files. This comparison narrows metadata races,
32
+ not the documented uncooperative-writer race. Tests simulate these SELinux probes;
33
+ they are not a live SELinux-host qualification.
34
+
35
+ On macOS, any extended-attribute (`@`) or ACL (`+`) marker deliberately refuses,
36
+ including `com.apple.provenance`. This v1 writer does not copy or remove those
37
+ attributes. Inspect the selected file; a refusal never authorizes stripping its
38
+ metadata automatically. Existing file permissions are preserved.
39
+ New config directories use 0700 and files 0600. The plan shows the effective file,
40
+ semantic key, current/desired trust scalar, byte digests and override provenance;
41
+ it never returns other config values or credentials.
42
+
43
+ This explicit act is the exception to read-only launch trust checks. Spawning,
44
+ scaffolding, starting and retiring do not write native trust or create consent
45
+ records. Existing Codex root-trust/yolo per-invocation overrides are unchanged.
46
+ This command does not enable development-channel prompt consent, change permission
47
+ mode, authenticate, install a plugin, or establish model/channel readiness.
48
+
49
+ ## Qualified scope and formats
50
+
51
+ Claude 2.1.289 darwin-arm64 was qualified against pinned native source and a
52
+ private no-Git deployment. One native root acceptance preceded two fresh child
53
+ launches. Root trust was true before each; neither child had accepted trust;
54
+ both reached ordinary UI. Native child bookkeeping later contained false.
55
+ Both bypass-mode controls in a separate disposable deployment still asked trust.
56
+ Native source independently walks trusted ancestors outside repository boundaries.
57
+ The recorded recipe and source exclude channel argv as a grant of trust. There
58
+ was no contemporaneous per-process environment snapshot; the same-runner root
59
+ trust question and absence of child recipe/capability bypass settings support
60
+ this bounded evidence. This is not qualification of other versions, nested Git
61
+ layouts or account/channel readiness.
62
+
63
+ Codex 0.160.1 native source independently declares a project `trust_level` and
64
+ writes `trusted`/`untrusted` under the project path. See the pinned
65
+ [native writer](https://github.com/openai/codex/blob/rust-v0.160.1/codex-rs/core/src/config/mod.rs)
66
+ and [configuration type](https://github.com/openai/codex/blob/rust-v0.160.1/codex-rs/config/src/config_toml.rs).
67
+ OATS's existing root-consent-to-exact-home invocation override remains separate
68
+ from native ancestor inheritance; Claude evidence is not Codex evidence.
69
+
70
+ JSON parsing rejects duplicate keys and wrong shapes. Only the required missing
71
+ object suffix or existing boolean token changes; unrelated large numbers, keys,
72
+ spacing and newlines remain byte-identical. False becomes true; an existing true
73
+ is a byte/mtime-preserving no-op.
74
+
75
+ The dependency-free TOML editor supports ordinary single-line assignments,
76
+ quoted/bare/dotted keys, explicit project tables, dotted trust keys and the
77
+ reader's inline project form containing only `trust_level`. It validates scalar,
78
+ single-line array and inline-table structure, duplicate/conflicting definitions,
79
+ and signed 64-bit integers. Multiline values, array-of-table headers, dates,
80
+ nondecimal numbers and unsupported representations refuse even if another TOML
81
+ parser accepts them. Root-level inline `projects` and inline project entries with
82
+ additional fields refuse rather than write a shape the existing readiness reader
83
+ does not consume. Comments, whitespace and line endings outside edit spans stay
84
+ unchanged. Untrusted becomes trusted; other trust values/types refuse. Both
85
+ editors validate the candidate and compare semantic trees with only the intended
86
+ leaf changed. Unsupported native format changes require refreshed qualification.
87
+
88
+ A Claude root entry cannot cover a home behind an intervening Git boundary.
89
+ Readiness may still warn after an unchanged root apply. Inspect that boundary;
90
+ do not automatically write a second root or answer a trust dialog.
91
+
92
+ ## Durability, audit and recovery
93
+
94
+ All selected configs are preflighted before any native write. Cooperating OATS
95
+ writers acquire exclusive destination locks in canonical-path order plus an audit
96
+ lock. Existing locks refuse promptly. The human/JSON error message names the exact
97
+ lock path so an operator can inspect an interrupted apply; OATS never steals or
98
+ deletes another invocation's lock, even when it appears stale. Identity and
99
+ raw-byte digests are rechecked after locking and immediately before replacement.
100
+ Candidate files are exclusively created beside the destination, fsynced, atomically
101
+ renamed and verified through the existing trust reader. Native writers do not
102
+ honor OATS locks: final checks detect observed drift but cannot guarantee exclusion
103
+ against an uncooperative writer racing the rename. No native process is stopped.
104
+
105
+ The deployment-private `.agents/harness-trust.jsonl` is independent of instance
106
+ lifetime, alongside other private kernel events. Its new directory/file modes are
107
+ 0700/0600. A checked, durable intent precedes each changed file; a checked outcome
108
+ follows it. An unchanged apply records an outcome without touching config bytes
109
+ or mtime. Plan creates no audit, directory, lock or temporary file. Invalid or
110
+ truncated existing JSONL refuses until operator inspection.
111
+
112
+ `all` is not a transaction. Failure stops later work, retains earlier changes and
113
+ reports unattempted entries. Trust may already have been consumed, so no rollback
114
+ is simulated. Inspect the native files and audit before rerunning; a correct
115
+ existing entry converges as a no-op. `E_HARNESS_TRUST_INCOMPLETE` has closed reasons
116
+ `locked|changed|audit|write|verify` and a conservative `mayHaveChanged` flag.
117
+ An audit failure before a write prevents it; failure after a possible replacement
118
+ reports incomplete and preserves intent evidence. The command never claims that
119
+ failure means native trust was undone.
120
+
121
+ ### Native qualification references
122
+
123
+ Claude source qualification used executable SHA256
124
+ `03d66745e3bb69ec727d66023696f3820bc0a00a8a5ba725eb6706d0c67cbe69`.
125
+ Its native exact/ancestor lookup is in byte range 182226330–182227750;
126
+ the startup trust decision is 197759244–197759315. The lookup has no
127
+ permission-mode or channel-argument input. Environment gates remain a separate
128
+ limit: `CLAUDE_CODE_SANDBOXED`, `IS_DEMO`, `CLAUBBIT` and the native `Le(false)`
129
+ gate can bypass the question. The disposable observation did not record a
130
+ contemporaneous child-process environment snapshot. No proprietary source is
131
+ included here.
132
+
133
+ Codex qualification is source-derived from `rust-v0.160.1`, independently of the
134
+ Claude observation. `set_project_trust_level_inner` (core config lines 2343–2410)
135
+ writes the exact project key and trust scalar; `ProjectConfig` (config type
136
+ lines 603–615) consumes the enum. The
137
+ [loader](https://github.com/openai/codex/blob/rust-v0.160.1/codex-rs/config/src/loader/mod.rs)
138
+ uses exact normalized project lookup keys (lines 1030–1082), preferring canonical
139
+ path spelling (1370–1398). This qualifies the native representation; OATS retains
140
+ its separately tested invocation override instead of claiming universal native
141
+ ancestor inheritance. The source files' SHA256 digests are:
142
+
143
+ | Upstream path | SHA256 |
144
+ | --- | --- |
145
+ | `codex-rs/core/src/config/mod.rs` | `854f5a6b184714740500b8d3eae2bb04cb88172efccf11d38a74f2f6189c0341` |
146
+ | `codex-rs/config/src/config_toml.rs` | `453b5ed4d49cf3768ce47e1f2fe1987c92d694bba9b52373dcf10f15901ab9d1` |
147
+ | `codex-rs/config/src/loader/mod.rs` | `2ec85a665f827dd7080f8e562c960e49158b0d8a11b1022cc0a93eed05986d0c` |
@@ -13,8 +13,9 @@ reference pages ([workspaces](workspaces.md), [souls and instances](souls-and-in
13
13
  catalog (`package-catalog.json`). It has no runtime dependencies.
14
14
  - **`@awebai/oats-pi`** (`packages/pi/`): a thin pi adapter that exposes an
15
15
  instance's own resources. It registers no agent tools.
16
- - **`oats.framework`** (`oats-package/`): the `oats.core`, `oats.setup` and
17
- `oats.knowledge-theory` capabilities and the `knowledge-theory-expert` soul,
16
+ - **`oats.framework`** (`oats-package/`): the `oats.core`, `oats.setup`,
17
+ `oats.support` and `oats.knowledge-theory` capabilities and the
18
+ `knowledge-theory-expert` soul,
18
19
  released as a package under its own `oats-framework/v<version>` tags.
19
20
  `oats.core` is the one with executables: its spawn and launch hook and the
20
21
  Claude Code waiting emitter (`capabilities/oats-core/bin/`).
@@ -53,7 +54,8 @@ published to npm. Its developer docs are in
53
54
  | `capability-show.mjs` | `oats capabilities show`: one catalog row's inject and skill files, read at its commit |
54
55
  | `materialize.mjs` | copying modules into a home and composing it |
55
56
  | `core.mjs` | spawn, retire, sessions, hooks, launch recipes, instance metadata |
56
- | `instruction-composition.mjs` | the generated `AGENTS.md` |
57
+ | `instruction-composition.mjs` | the generated `AGENTS.md`; `renderInstructionParts` is the span-reporting renderer (`renderInstructionText` is its text) |
58
+ | `soul-composition.mjs` | the instructions a spawn of a soul would write, composed without a spawn (kernel half + materialize into a scratch home): the one path behind `doctor --soul` and `inspect --soul --instructions` |
57
59
  | `teams.mjs`, `teams-verbs.mjs` | the team model and the `oats teams` verbs |
58
60
  | `schedule.mjs`, `schedule-host.mjs`, `triggers.mjs`, `automations.mjs` | schedules, triggers and the host timer |
59
61
  | `schedule-command.mjs`, `schedule-command-child.mjs` | synchronous scheduler adapter and asynchronous child supervisor: bounded output, TERM/KILL escalation and observed exit; no changes to synchronous tick/lock callbacks |
@@ -65,9 +67,13 @@ published to npm. Its developer docs are in
65
67
  | `capability-contract.mjs`, `provider-binding.mjs` | manifest validation (launch environment, hooks, `retirement`), the hook environment rules, the readiness wire |
66
68
  | `retire-output.mjs` | the lines `oats retire` prints for preserved work, one function for the local and the remote path |
67
69
  | `servers.mjs` | routing commands to a registered server |
70
+ | `launch-prompt-evidence.mjs` | private bounded storage of exact post-answer unmatched or structurally completed frames, original-home identity checks and receipt-last publication; no terminal operations or public raw text |
71
+ | `launch-prompt-completion.mjs` | pure, qualified post-answer completion reporting using a finite bottom input/footer table and 21 question markers; never an input matcher |
68
72
  | `launch-prompt-policy.mjs` | strict host-local exact-home aweb development-channel consent, canonical directory validation and config provenance |
69
73
  | `launch-prompts.mjs`, `launch-prompt-transport.mjs` | bounded launch-only controller and visible-screen transport; private process authority, injectable clock/transport, durable checked receipts and the fixture-gated exact aweb development-channel signature |
70
- | `harness-trust.mjs` | reading (never writing) Claude's and Codex's folder trust for a launch |
74
+ | `harness-trust.mjs` | read-only native trust coverage and operator remedy for launches |
75
+ | `harness-trust-write.mjs` | explicit operator plan/apply, cooperating locks, durable deployment audit and no-rollback outcomes |
76
+ | `harness-trust-json.mjs`, `harness-trust-toml.mjs` | bounded native parsers and byte-preserving edits of one trust leaf |
71
77
 
72
78
  The schedule registry stores explicit concurrency caps, leaving the schedule
73
79
  cap absent for its effective default of five. `readRegistry()` is a pure,
@@ -0,0 +1,70 @@
1
+ # Qualified post-answer completion
2
+
3
+ ## Boundary and accepted risk
4
+
5
+ Only after exactly one recorded **submitted** development-channel Enter, under the existing qualified executable/version/platform/argv, exact-home consent, original home identity, pane/PID and 110x35 checks, may structural recognition complete observation. It never authorizes input. Existing exact completion families remain available, subject to the marker precedence below. No structural recognition before an answer, after any terminal outcome, or after failed/uncertain input.
6
+
7
+ An unknown dialog can retain the composer and footer and evade this finite marker list. The contract explicitly accepts that completion-reporting risk. Completion closes input authority permanently; it does not establish absence of every dialog, idle state, account admission, channel readiness, or task success. The observed spinner is an in-progress task turn, not an idle prompt. No spinner word, glyph or timing grammar is introduced.
8
+
9
+ ## Source provenance
10
+
11
+ Read-only pinned Claude Code 2.1.289 darwin-arm64 binary SHA256 `03d66745e3bb69ec727d66023696f3820bc0a00a8a5ba725eb6706d0c67cbe69`. [the source manifest](../lib/launch-prompt-fixtures/structural-completion.json) records exact byte intervals and SHA256s for every reference below. No extracted module enters the repository. These are source-derived rules, not additional observed captures.
12
+
13
+ ## Qualified input/footer rule — non-flag footer layout
14
+
15
+ This qualifies the non-Jt footer layout (`tengu_copper_thistle` false) in the pinned source. The flag-enabled layout chooses a single hint instead of this list; it is not separately qualified. A remote flag change can alter the footer without changing the binary and reduce matches; only exact listed shapes match, never additional shapes.
16
+
17
+ Input is the same compared text object, exactly 35 newline-terminated rows at 110x35, bounded to 64 KiB. No whitespace normalization, cropping, ANSI stripping, case folding or transcript rewrite. Reject embedded CR, ESC and other C0 controls except LF. All rows after the footer must be empty. The footer must be the last nonempty row, with this immediately preceding contiguous region:
18
+
19
+ 1. Exactly 110 U+2500 horizontal-line characters.
20
+ 2. Exactly U+276F followed by U+00A0 (`❯` plus NBSP), and nothing else: the actual empty composer row, not a prompt-like substring in prose.
21
+ 3. Exactly 110 U+2500 horizontal-line characters.
22
+ 4. The finite footer below.
23
+
24
+ This accepts the region at any row where all four rows fit, not a fixed transcript height. It refuses nonempty/multiline editor drafts, shell or screen-reader selectors, annotated borders, wrapped footers, notifications between editor and footer, status rows below the footer, and footer-less states. It deliberately does not cover every normal renderer configuration.
25
+
26
+ Source: `prompt-selector` s0/qK renders Z.pointer+NBSP for the ordinary non-screen-reader input; `prompt-symbol` pins Z.pointer to U+276F. `round-border-glyphs` pins top/bottom to U+2500; `prompt-border` BA and `editor-layout` HWe give top/bottom borders with left/right borders absent. The existing raw API captures and the actual retained Max frame corroborate the exact three input rows at 110 columns. `footer-layout` YHe/Oie gives paddingX=2, row width=110, gap=1, right status marginLeft=auto. The sanitized regression frame's input region is rows 11–14; its home-only sanitized derivative is regression input, not an exact spinner matcher.
27
+
28
+ ### Modes (complete list)
29
+
30
+ | Native mode | Exact left label | Allowed cycle suffix |
31
+ |---|---|---|
32
+ | default | `⏸ manual mode on` | absent only |
33
+ | plan | `⏸ plan mode on` | absent or ` (shift+tab to cycle)` |
34
+ | acceptEdits | `⏵⏵ accept edits on` | absent or same cycle suffix |
35
+ | bypassPermissions | `⏵⏵ bypass permissions on` | absent or same cycle suffix |
36
+ | dontAsk | `⏵⏵ don't ask on` | absent or same cycle suffix |
37
+ | auto | `⏵⏵ auto mode on` | absent or same cycle suffix |
38
+
39
+ `mode-map` Sr/DL/bie/qQo enumerates all six; `mode-symbols` pins the symbols; `footer-mode` kx adds literal ` on` when bare is absent. `footer-branches` dHe calls kx without bare and gates the default-keybinding cycle hint on non-default mode and ks. Unknown modes, custom keybindings and bare labels are unsupported.
40
+
41
+ After the mode/cycle label, accept only these exact suffixes, each independently enumerated: empty; ` · ← for agents`; ` · esc to interrupt`; ` · esc to interrupt · ← for agents`. For **default mode only**, also ` · ? for shortcuts` or ` · ? for shortcuts · ← for agents`. Sources: `footer-branches` dHe/EFt (escape interrupt, shortcut hint, ordering), `agent-hint` Cze (left-arrow for agents). No return-to-team-lead, task counter, voice, custom status or arbitrary suffix. The normal hints are not negative markers anywhere in the visible frame; an unsupported footer remains unsupported for structural reasons rather than because it contains a question mark or Esc.
42
+
43
+ The footer is exactly two ASCII spaces + left text + ASCII space gap + one right label from this complete list: `○ low · /effort`, `◐ medium · /effort`, `● high · /effort`, `◉ xhigh · /effort`, `◈ max · /effort`. `effort` bW/y8 and `mode-symbols` define these glyphs and labels. The right label ends at column 108 (two columns right padding are absent in tmux text); gap is exactly `108 - 2 - leftDisplayWidth - rightDisplayWidth`, and at least one. All enumerated glyphs are single-column in the qualified captured geometry. No `effort:` prefix, ultracode, missing effort, altered punctuation, or other effort value. Existing medium-only exact families remain exact; the additional finite effort rules are source-qualified here, not prior capture qualification.
44
+
45
+ ## Qualified negative marker table
46
+
47
+ ONLY after exactly one recorded submitted answer, each listed string is matched **case-sensitively as exact contiguous text anywhere in the visible frame**, before either exact or structural positive completion. It can conservatively block a transcript quoting a known question. No regex/question-word heuristic, case folding, line joining, general `?`, general `Esc`, or exemption for an unknown question. Wrapped or customized unknown questions remain within the explicitly accepted risk. Input prompt matching itself is unchanged: the initial qualified channel question must still be answerable before a submitted answer exists.
48
+
49
+ | Exact forms (each independently tested) | Pinned source reference |
50
+ |---|---|
51
+ | `Accessing workspace:`; `Quick safety check:`; `Yes, I trust this folder`; `No, continue without these permissions`; `No, exit` | trust / TrustDialog |
52
+ | `WARNING: Loading development channels`; `I am using this for local development` | channel confirmation |
53
+ | `Managed settings require approval`; `Yes, I trust these settings`; `No, exit Claude Code` | managed-settings |
54
+ | `Do you want to use this API key?`; `Detected a custom API key in your environment` | api-key |
55
+ | `Select login method:` | login |
56
+ | `Do you want to proceed?` | permission-question / o5 default question |
57
+ | `Do you want to allow this connection?` | connection-question |
58
+ | `Waiting for permission…` | permission-wait / QVe |
59
+ | `Enter to confirm`; `Esc to cancel` | confirmation-guide and key-hint B renderer; gateway/selection-dialog corroborate cancel literal |
60
+ | `esc to cancel` | lower-cancel + key-hint B with lower keyCase |
61
+ | `Enter to select` | selection-dialog literal |
62
+ | `Esc to exit` | managed-settings + key-hint B |
63
+
64
+ The historical local-JSX confirmation is not claimed covered by this list; its marker composition was not separately verified and remains an example of the accepted unknown-dialog risk. Real #712 child screens with missing effort and update/focus notices below the footer remain unqualified; the retained A2 frame has the required effort and no trailing notices.
65
+
66
+ This is the whole v1 marker set, **not a claim that these are all possible native questions**. Generic custom title/confirm labels can evade it. `? for shortcuts` and `esc to interrupt` are explicitly normal UI, never markers. No case-insensitive expansion to `enter to confirm`, `esc to exit`, etc. without another qualified table update.
67
+
68
+
69
+ Private retention and receipt-v2 collection rules are documented in
70
+ [execution targets](execution-targets.md#private-unmatched-launch-evidence).
@@ -9,7 +9,7 @@ or workspace membership alone does not make a package official.
9
9
 
10
10
  | package | release | capabilities | package souls |
11
11
  |---|---|---|---|
12
- | `oats.framework` | `oats-framework/v1.6.1` (this repository) | `oats.core`, `oats.setup`, `oats.knowledge-theory` | `knowledge-theory-expert` |
12
+ | `oats.framework` | `oats-framework/v1.7.0` (this repository) | `oats.core`, `oats.setup`, `oats.knowledge-theory` | `knowledge-theory-expert` |
13
13
  | `oats.okf` | `v4.1.1` | `oats.okf` (knowledge), `oats.okf-harvest`, `oats.okf-maintenance` | `knowledge-harvester`, `knowledge-maintainer` |
14
14
  | `oats.aweb` | `v1.21.1` | `oats.aweb` (messaging) | |
15
15
  | `oats.engineering` | `v1.8.1` | `oats.engineering-expert`, `oats.developer`, `oats.code-review`, `oats.maintainer` | `code-reviewer` |
package/docs/packages.md CHANGED
@@ -51,7 +51,7 @@ packages:
51
51
  - **Bare version** (`v4.1.1`, `4.1.1`, `1.0.0-rc.1`): the id is looked up in
52
52
  the official catalog — `package-catalog.json` in the `oats` repo, or the file
53
53
  named by `OATS_PACKAGE_CATALOG` — which supplies the repo url, the tag
54
- convention (`v4.1.1` or `oats-framework/v1.6.1`) and the payload path. An id
54
+ convention (`v4.1.1` or `oats-framework/v1.7.0`) and the payload path. An id
55
55
  the catalog does not know is `E_PACKAGE_MISSING` ("use `git:<repo>@<ref>` for
56
56
  a package outside the catalog"). The catalog is the reviewed official list
57
57
  ([official-catalog.md](official-catalog.md)) and the only way a
@@ -74,7 +74,7 @@ members:
74
74
  - git:github.com/acme/agents
75
75
  - git:github.com/acme/platform
76
76
  packages:
77
- oats.framework: v1.6.1
77
+ oats.framework: v1.7.0
78
78
  oats.okf: v4.1.1
79
79
  oats.aweb: v1.21.1
80
80
  teams:
@@ -339,11 +339,11 @@ A soul that names one of the package's capabilities with
339
339
  "policy": "docs/official-catalog.md",
340
340
  "packages": {
341
341
  "oats.okf": { "url": "https://github.com/awebai/oats-okf.git", "ref": "v4.1.1", "path": "oats-package" },
342
- "oats.framework": { "url": "https://github.com/awebai/oats.git", "ref": "oats-framework/v1.6.1", "path": "oats-package" }
342
+ "oats.framework": { "url": "https://github.com/awebai/oats.git", "ref": "oats-framework/v1.7.0", "path": "oats-package" }
343
343
  }
344
344
  }
345
345
  ```
346
346
 
347
- `ref` carries the tag convention: a workspace's `oats.framework: v1.6.1`
348
- resolves to tag `oats-framework/v1.6.1`. Resolving through the catalog never
347
+ `ref` carries the tag convention: a workspace's `oats.framework: v1.7.0`
348
+ resolves to tag `oats-framework/v1.7.0`. Resolving through the catalog never
349
349
  advances a lock by itself: `oats sync` does, and says so.
@@ -0,0 +1,71 @@
1
+ # OATS 0.46.0
2
+
3
+ ## Added
4
+
5
+ - **`oats inspect --soul <soul> --instructions` shows the instructions a new
6
+ instance would get.** `souls[0].composedInstructions` is the AGENTS.md a
7
+ spawn of that soul here would write: the soul's own text, then the
8
+ instance-boundary and work-mode blocks and each capability's inject,
9
+ produced by spawn's own composition, with nothing written to the
10
+ deployment. The soul's text and each block are given as ranges into the
11
+ text, named like an instance's `instructions.sources`, so the Desktop can
12
+ show which part comes from where. Without `--instructions` nothing extra is
13
+ composed; with `--home` it is refused. Feature `soul-composed-instructions`.
14
+ `oats doctor --soul` uses the same composition, and its capability markers
15
+ now name the inject home-relative instead of a deleted temporary path.
16
+ Routed with `--server <id>`, a host whose probe does not advertise
17
+ `soul-composed-instructions` is refused with `E_REMOTE_INCOMPATIBLE`, and
18
+ nothing is sent.
19
+
20
+ ## Fixed
21
+
22
+ - **Interim subscription completion extension:** the qualified Claude 2.1.289
23
+ darwin-arm64 110x35 empty Opus 5.5 / auto / medium frames additionally accept
24
+ exactly Claude Max, Pro, Team and Enterprise billing labels. These variants
25
+ are source-derived synthetic frames, not captured subscription sessions.
26
+ The exact Enter-authorizing prompt, explicit home consent and all launch
27
+ guards are unchanged. Completion closes input authority and proves no
28
+ account/channel readiness. This is **not the #754 onboarding fix**: #754
29
+ remains open for task-filled and bypass-permissions completion recognition.
30
+ Other unqualified frames still retain a blocked result after a submitted
31
+ answer; there is no automatic retry.
32
+
33
+ - **`oats doctor --soul` lists every inject block of a capability that
34
+ declares several.** Its `instructionBlocks` kept only the first block of
35
+ such a capability, though its composed text had them all; it now lists
36
+ each one, as a spawn records them.
37
+
38
+ ## Changed
39
+
40
+ - **The Desktop accepts OATS CLIs `>=0.25.8 <0.47.0`.** Its kernel band
41
+ admits 0.46; the soul composed-instructions feature is not read by the
42
+ Desktop yet, so no new gate is needed.
43
+
44
+ - **oats.framework 1.7.0: oats.setup 2.4.0 and oats.core 2.4.2.** Instruction
45
+ text only; no kernel, schema or provider contract changes.
46
+ - **oats.setup 2.4.0: team setup instructions are executable from one
47
+ authorized intake.** The onboarding and teams skills carry one
48
+ authorized intake through team configuration, operator composition
49
+ checks and first-task verification; they link the provider-owned
50
+ identity and delivery procedures, distinguish unsupported paths, and keep
51
+ project operator advice independent of central knowledge identities. An
52
+ instruction change, not a new team or provider primitive
53
+ ([#709](https://github.com/awebai/oats/issues/709), #721). The setup
54
+ inject and `oats-automations` keep configuration repairs proportional to
55
+ the requested outcome (#689, #692) and describe schedule and trigger
56
+ descriptions, retained-attempt reconciliation and supervisor cleanup as
57
+ the kernel now does them (#684, #591, #592, #593).
58
+ - **oats.core 2.4.2:** `oats-operate` covers the extra work trees a
59
+ work-mode briefing allows and their retirement (#678), and what
60
+ `oats session input` does and does not prove (#595).
61
+
62
+ - Add explicit host-local `oats harness trust --plan` and bare apply for one
63
+ canonical deployment-root Claude/Codex trust entry. Edits preserve unrelated
64
+ native bytes and record a durable `.agents/harness-trust.jsonl` audit. Partial
65
+ failures are retained with `E_HARNESS_TRUST_INCOMPLETE`; no rollback is claimed.
66
+ Spawn/start/retire do not write native trust. Claude ancestor qualification is
67
+ bounded to2.1.289 darwin-arm64/no-Git paths; Codex format qualification is separate.
68
+ Nested Git boundaries and unsupported native formats still refuse or warn.
69
+ Native canonicalization preserves on-disk root casing on APFS; result keys,
70
+ audit and reader verification agree. Lock refusals name the exact blocking
71
+ lock in human/JSON errors without stealing or deleting it.
@@ -0,0 +1,93 @@
1
+ # OATS 0.47.0
2
+
3
+ 2026-10-07
4
+
5
+ ## Changed
6
+
7
+ - **The Desktop accepts OATS CLIs `>=0.25.8 <0.48.0`.** The compatibility band
8
+ admits 0.47; existing feature gates remain unchanged.
9
+
10
+ - **Desktop: "Retire instance" says what will happen, and shows it happening.**
11
+ The roster row's menu item is **Retire instance…** (was "Remove instance…"),
12
+ and the confirmation reads **Retire <name>?** with a **Retire instance**
13
+ button. Stop shares the same dialog: **Stop <name>?**, **Stop session**
14
+ (**Stop sessions** for several).
15
+ - **Plain words.** **What will happen** lists the consequences from the
16
+ plan: the session stops, children stop first and are kept, a recovery copy
17
+ of uncommitted work is saved before the home folder is deleted, the
18
+ worktree is kept on its branch or deleted, and branches and pull requests
19
+ are not changed. **Current state** shows Running / Not running / Unknown,
20
+ the uncommitted work and the branch. The plan's time is shown as an age
21
+ ("Observed just now"), with no timestamps or kernel vocabulary.
22
+ - **Loading and progress.** The dialog opens with a skeleton and "Checking
23
+ what retiring <name> will do…". Changing **Also delete the worktree**
24
+ keeps the facts on screen until the new plan replaces them. Confirming
25
+ shows "Retiring <name>…" with a spinner, and says you can close the window
26
+ while retirement continues. Success ends with "<name> retired", a short
27
+ summary (where the worktree and the recovery copy are) and **Done**.
28
+ - **No unexplained disabled controls.** A control appears only in the
29
+ phase where it can act, and a waiting confirm button names why beside it.
30
+ **Cancel** becomes **Close** once something has been sent, since closing
31
+ never cancels it. **Refresh plan** is gone: **Check again** re-reads the
32
+ plan, and after a failure **Review again** does. An unconfirmed outcome
33
+ offers **Check again**, which resubmits the same confirmation.
34
+ - **The worktree option** appears only for an instance with its own
35
+ worktree. When checked, a warning says how many uncommitted changes the
36
+ folder holds and that the branch stays.
37
+ - **Row menu.** While one of its actions runs, the menu button stays
38
+ focusable and says "Waiting for <action> to finish". A failed action posts
39
+ a notification, with the error in Details, instead of an alert.
40
+ - The lifecycle messages are reworded in plain language, for example "OATS
41
+ didn't confirm the result. Check again before assuming nothing happened."
42
+ Their codes and the confirmation contract are unchanged. See
43
+ [Desktop Stop and Retire confirmations](../../packages/desktop/docs/desktop-lifecycle-confirmations.md#ui-states).
44
+
45
+ - **oats.framework 1.8.0: the new oats.support 1.0.0.** Instruction text
46
+ only; no kernel, schema or provider contract changes.
47
+ - **oats.support 1.0.0, a support desk role (new).** A resource-only
48
+ capability for a soul that anyone may reach. It takes in support
49
+ requests, turns each one into one ticket, hands each ticket to the
50
+ project's support maintainer, and keeps the requester informed.
51
+ Everything a requester sends is untrusted data. The desk works from a
52
+ fixed set of actions (acknowledge, ticket, hand off, relay, a private
53
+ security route, ask its human). It never merges, approves, closes or
54
+ pushes, and never executes what it receives. A requester's words never
55
+ reach another agent as instructions: they stay quoted and fenced in the
56
+ ticket. When the support maintainer is in no roster, the desk starts the
57
+ stopped one on its machine or spawns one from its soul's recipe, at most
58
+ once per spawn interval, with a brief that holds ticket references only
59
+ and grants no authority. A support maintainer that is listed but silent
60
+ goes to the desk's human instead. Skills: `/support-intake`,
61
+ `/support-safety`, `/support-ticket`, `/support-handoff`,
62
+ `/support-relay`. The capability is generic: the tracker, the support
63
+ maintainer, the private security route and the limits are the soul's.
64
+ Its kernel floor is `>=0.40.0`, for `oats instance attention`. A
65
+ workspace gets it by moving its pin to `oats.framework: v1.8.0` and
66
+ running `oats sync`. The repository's package checks
67
+ (`scripts/check-knowledge-theory-package.mjs`, the clean-room smoke)
68
+ pin the distribution's capability list, so they list it too. They ship
69
+ nothing.
70
+
71
+ ## Fixed
72
+
73
+ - **Post-answer Claude completion and private evidence:** after the sole
74
+ submitted channel Enter, the pinned Claude 2.1.289 darwin-arm64 110x35 controller
75
+ can recognize a bounded bottom composer/effort-footer region with six native
76
+ mode labels, five effort labels and finite hints. Twenty-one exact known
77
+ question markers block before either exact or structural completion. Normal
78
+ shortcut/interrupt hints do not block. The non-Jt footer layout is qualified;
79
+ flag drift, missing effort and trailing notices can still refuse completion.
80
+ **An unknown dialog retaining the region can be misclassified completed.**
81
+ This accepted reporting risk cannot authorize another key or prove idle,
82
+ account or channel readiness. Existing Enter and consent/ownership guards are
83
+ unchanged. A home-only sanitized derivative of an actual in-progress Max
84
+ comparison is the regression fixture; there is no exact spinner matcher or
85
+ new live proof of the revised classifier. Also adds the source-qualified
86
+ empty Max/bypass/medium families.
87
+ Exact compared bytes are retained privately for structural completion and
88
+ post-answer unmatched blocks. Receipt v2 distinguishes those two outcomes;
89
+ historical v1 remains readable. A structural save failure reports checked
90
+ incomplete with answer/target preserved. No raw text enters public output.
91
+ See [completion scope](../launch-prompt-completion.md) and
92
+ [private evidence](../execution-targets.md#private-unmatched-launch-evidence).
93
+ This does not close #754 or establish general onboarding success.
package/docs/servers.md CHANGED
@@ -216,7 +216,7 @@ sent as text or on stdin, never as paths.
216
216
  | `okf harvest --instance <name>` | the saved home | `harvest` |
217
217
  | `schedule ...` | registered workspace | `schedule` |
218
218
  | `launch-config list\|set\|remove\|preview` | `--dir`, the instance's home, or the registered workspace | `launch-config` |
219
- | `inspect`, `operation run` | `--dir`, the instance's home, or the registered workspace | `operations` |
219
+ | `inspect`, `operation run` | `--dir`, the instance's home, or the registered workspace | `operations`; `inspect --instructions` also `soul-composed-instructions` |
220
220
  | `readiness` | the instance's home, `--dir`, or the registered workspace | `readiness` (`readinessApi: 2`) |
221
221
  | `instance events` | the instance's home | `instance-events-2` (`eventsApi: 2`) |
222
222
  | `instance git`, `instance diff` | the instance's home; the Git runs there | `instance-git` (`instanceGitApi: 1`) |
@@ -102,10 +102,15 @@ souls) is ordinary capability content: **`oats.core`** (package
102
102
  `oats.framework`). Workspaces give it to every soul through
103
103
  `defaults.capabilities: { oats.core: { from: package } }`; a soul may say
104
104
  `oats.core: off`. **`oats.setup`** (same package) carries the whole-architecture
105
- knowledge an onboarding expert needs. Neither is kernel magic; the kernel still
105
+ knowledge an onboarding expert needs, and **`oats.support`** (same package)
106
+ is the support desk role: a publicly reachable soul that tickets requests and
107
+ hands them to a support maintainer. None of them is kernel magic; the kernel still
106
108
  composes its own instance-boundary and work-mode briefings, and the "You run
107
109
  on OATS" briefing is `oats.core`'s inject: a soul without `oats.core` gets no
108
- OATS operating instructions, and `oats doctor --soul` says so.
110
+ OATS operating instructions, and `oats doctor --soul` says so. Its capability
111
+ markers' `src=` is home-relative (`.oats/modules/<cap>/<inject>`), composed by
112
+ the same path as `oats inspect --soul --instructions`; kernel markers keep the
113
+ installed package's path.
109
114
 
110
115
  ## Instance anatomy
111
116
 
@@ -210,18 +215,18 @@ instructions and pins model/provider settings; it excludes nothing.
210
215
  Claude Code and Codex ask before they work in a folder they have not seen, and
211
216
  every instance home is new. A launch stopped at that prompt waits for a human,
212
217
  so the operator trusts the **deployment directory** (where `oats-local.yaml`
213
- is) once per harness. OATS only reads the harnesses' configuration; it never
214
- writes it.
218
+ is) once per harness. Launches only read native configuration. The explicit
219
+ `oats harness trust --dir D --harness all --plan` previews the entry; the same
220
+ command without `--plan` applies it with a deployment audit. See
221
+ [native trust](harness-trust.md) for safe operation and partial-failure recovery.
215
222
 
216
223
  - **Claude Code** looks for an accepted entry for its folder or an ancestor, up
217
- to a git root. Instance homes are not inside a git repository, so one entry
218
- for the deployment covers every home under it. To add it, run `claude` in the
219
- deployment once and accept the prompt. That records
224
+ to a Git root. An intervening repository can prevent deployment-root inheritance;
225
+ the command does not automatically trust another root. The native field is
220
226
  `projects["<deployment>"].hasTrustDialogAccepted` in `~/.claude.json`
221
227
  (`$CLAUDE_CONFIG_DIR/.claude.json` when that is set).
222
228
  - **Codex** applies only an exact entry: a trusted parent does not cover the
223
- folders below it. To give the operator's consent, run `codex` in the
224
- deployment once and choose "Trust and continue". That records
229
+ folders below it. The explicit deployment-root command records
225
230
  `[projects."<deployment>"] trust_level = "trusted"` in
226
231
  `~/.codex/config.toml` (`$CODEX_HOME/config.toml`). With that entry, or one
227
232
  for an ancestor of the deployment, each codex launch trusts its own new home
@@ -54,7 +54,7 @@ members: # repo refs, NO @revision (E_WORKSPAC
54
54
  - git:github.com/acme/tools # a member that ALSO publishes a package (see below)
55
55
 
56
56
  packages: # the ONLY versioned things
57
- oats.framework: v1.6.1 # bare version → resolves through the official catalog
57
+ oats.framework: v1.7.0 # bare version → resolves through the official catalog
58
58
  oats.okf: v4.1.1
59
59
  acme.tools: git:github.com/acme/tools@v0.4.0 # outside the catalog → git:<repo>@<tag|OID>; still a package
60
60
 
package/lib/core.mjs CHANGED
@@ -29,6 +29,7 @@
29
29
  */
30
30
  import { effectiveLaunchPromptPolicy } from "./launch-prompt-policy.mjs";
31
31
  import { createLaunchPromptController } from "./launch-prompts.mjs";
32
+ import { retainLaunchPromptFrame } from "./launch-prompt-evidence.mjs";
32
33
  import { qualifyLaunchPromptFixtures } from "./launch-prompt-fixtures.mjs";
33
34
  import { LAUNCH_PANE_FORMAT, launchPaneTarget, launchPromptTransport } from "./launch-prompt-transport.mjs";
34
35
  import { execFileSync, execSync, spawn as spawnProcess, spawnSync } from "node:child_process";
@@ -646,7 +647,7 @@ function promptController(home, startId, harness, policy, command, io, expectedH
646
647
  const described = describeLaunchCommand(command);
647
648
  const qualification = qualifyLaunchPromptFixtures({ home: realPathOrNearest(home), harness, executablePath: described.executable, argv: described.argv });
648
649
  return createLaunchPromptController({ home: realPathOrNearest(home), startId, ...qualification, policy, expectedHomeIdentity,
649
- transport: launchPromptTransport(io), audit: data => appendEvent(home, { kind: "launch-prompt", data }),
650
+ transport: launchPromptTransport(io), audit: data => appendEvent(home, { kind: "launch-prompt", data }), retainFrame: retainLaunchPromptFrame,
650
651
  ...(io?.now ? { now: io.now } : {}), ...(io?.sleep ? { sleep: io.sleep } : {}) });
651
652
  }
652
653
  function retainedPromptError(home, instance, target, parentLineageCommitted, launchPrompts) {
@@ -0,0 +1,81 @@
1
+ /** Lossless JSON span edits for the one native Claude trust leaf. The parser
2
+ * rejects duplicate keys and never converts unrelated numeric tokens to Number. */
3
+ const bad = () => { throw new Error('unsupported or invalid native JSON configuration'); };
4
+ export function jsonDocument(text) {
5
+ let i = 0;
6
+ const space = () => { while (/\s/.test(text[i] ?? '') && i < text.length) { if (!' \r\n\t'.includes(text[i])) bad(); i++; } };
7
+ const string = () => {
8
+ const start = i++;
9
+ while (i < text.length) {
10
+ if (text[i] === '\\') { i += 2; continue; }
11
+ if (text[i++] === '"') { try { return { start, end: i, value: JSON.parse(text.slice(start, i)) }; } catch { bad(); } }
12
+ }
13
+ bad();
14
+ };
15
+ const value = (depth = 0) => {
16
+ if (depth > 128) bad();
17
+ space(); const start = i;
18
+ if (text[i] === '{') {
19
+ i++; space(); const fields = new Map();
20
+ if (text[i] !== '}') for (;;) {
21
+ if (text[i] !== '"') bad();
22
+ const key = string(); if (fields.has(key.value)) bad();
23
+ space(); if (text[i++] !== ':') bad();
24
+ fields.set(key.value, value(depth + 1)); space();
25
+ if (text[i] !== ',') break;
26
+ i++; space();
27
+ }
28
+ if (text[i++] !== '}') bad();
29
+ return { kind: 'object', start, end: i, fields };
30
+ }
31
+ if (text[i] === '[') {
32
+ i++; space(); const items = [];
33
+ if (text[i] !== ']') for (;;) {
34
+ items.push(value(depth + 1)); space();
35
+ if (text[i] !== ',') break;
36
+ i++; space();
37
+ }
38
+ if (text[i++] !== ']') bad();
39
+ return { kind: 'array', start, end: i, items };
40
+ }
41
+ if (text[i] === '"') return { kind: 'string', ...string() };
42
+ const match = /^(?:true|false|null|-?(?:0|[1-9][0-9]*)(?:\.[0-9]+)?(?:[eE][+-]?[0-9]+)?)/.exec(text.slice(i));
43
+ if (!match) bad();
44
+ i += match[0].length;
45
+ return { kind: 'scalar', start, end: i, token: match[0] };
46
+ };
47
+ const node = value(); space(); if (i !== text.length) bad(); return node;
48
+ }
49
+ const fingerprint = (n) => n.kind === 'object' ? ['object', [...n.fields].map(([k, v]) => [k, fingerprint(v)])]
50
+ : n.kind === 'array' ? ['array', n.items.map(fingerprint)] : [n.kind, n.value ?? n.token];
51
+ /** Return the exact candidate bytes; no configuration I/O or host defaults. */
52
+ export function editClaudeTrust(text, root) {
53
+ const document = jsonDocument(text);
54
+ const path = ['projects', root, 'hasTrustDialogAccepted'];
55
+ let parent = document;
56
+ for (let at = 0; at < path.length; at++) {
57
+ if (parent.kind !== 'object') bad();
58
+ const key = path[at], child = parent.fields.get(key);
59
+ if (!child) {
60
+ let inserted = true;
61
+ for (let n = path.length - 1; n > at; n--) inserted = { [path[n]]: inserted };
62
+ const insertion = `${parent.fields.size ? ',' : ''}${JSON.stringify(key)}:${JSON.stringify(inserted)}`;
63
+ const candidate = text.slice(0, parent.end - 1) + insertion + text.slice(parent.end - 1);
64
+ const after = jsonDocument(candidate);
65
+ // Independently derive the expected tree: only the missing suffix may appear.
66
+ parent.fields.set(key, jsonDocument(JSON.stringify(inserted)));
67
+ if (JSON.stringify(fingerprint(document)) !== JSON.stringify(fingerprint(after))) bad();
68
+ return { text: candidate, current: null };
69
+ }
70
+ if (at === path.length - 1) {
71
+ if (child.kind !== 'scalar' || !['true', 'false'].includes(child.token)) bad();
72
+ if (child.token === 'true') return { text, current: true };
73
+ const candidate = text.slice(0, child.start) + 'true' + text.slice(child.end);
74
+ child.token = 'true';
75
+ if (JSON.stringify(fingerprint(document)) !== JSON.stringify(fingerprint(jsonDocument(candidate)))) bad();
76
+ return { text: candidate, current: false };
77
+ }
78
+ parent = child;
79
+ }
80
+ bad();
81
+ }