docks-kit 0.22.25 → 0.22.28

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.
package/AGENTS.md CHANGED
@@ -69,7 +69,7 @@ Module map for `cli/` (engine axes, commands, kit store, test fixtures, and the
69
69
 
70
70
  Codex SoT notes:
71
71
  - `SoT/.codex/AGENTS.md` deploys to `~/.codex/AGENTS.md` as global Codex instructions.
72
- - `SoT/.codex/config.toml` pins Codex to `model = "gpt-6.1-sol"`, sets normal and plan reasoning to `high` with concise summaries, and sets `model_verbosity = "low"`, live top-level `web_search`, workspace-write sandboxing with sandboxed command network access, cross-session `memories` (+ dedicated note tools), `[agents]` subagent limits (`max_threads = 12`, `max_depth = 2` — intentionally above Codex defaults for broad parallel kit work; deeper recursion increases cost and predictability risk), a 128 KiB `project_doc_max_bytes` budget for the repo-side AGENTS.md chain (the global `~/.codex/AGENTS.md` is uncapped and not counted), and enables the two Docks plugins `docks@docks` and `plan-lifecycle@docks` (the shared plan lifecycle).
72
+ - `SoT/.codex/config.toml` pins Codex to `model = "gpt-6.2-sol"`, sets normal and plan reasoning to `high` with concise summaries, and sets `model_verbosity = "low"`, live top-level `web_search`, workspace-write sandboxing with sandboxed command network access, cross-session `memories` (+ dedicated note tools), `[agents]` subagent limits (`max_threads = 12`, `max_depth = 2` — intentionally above Codex defaults for broad parallel kit work; deeper recursion increases cost and predictability risk), a 128 KiB `project_doc_max_bytes` budget for the repo-side AGENTS.md chain (the global `~/.codex/AGENTS.md` is uncapped and not counted), and enables the two Docks plugins `docks@docks` and `plan-lifecycle@docks` (the shared plan lifecycle). The two project Codex reviewer agents in `.codex/agents/` use the same model.
73
73
  - `SoT/.codex/config.toml` does not set `personality`. Codex 0.159.0 deprecated the key, and `friendly` and `pragmatic` no longer select a style. `codexConfig.ts removeRetiredTopLevelSettings, retired top-level key prune` removes the retired `personality = "pragmatic"` line from a deployed `config.toml` on every sync. A different user-set value survives.
74
74
  - `SoT/.codex/rules/*.rules` deploys to `~/.codex/rules/` as kit-managed Codex command policy. This is Codex's equivalent of permission allow/prompt/block rules; user-learned approvals in `~/.codex/rules/default.rules` are preserved. Codex applies the strictest matching rule across all files, so a user allow rule cannot loosen a kit `forbidden` or `prompt` rule. Because the strictest rule wins, a narrow allow cannot carve an exception out of a broader prompt prefix, and a prefix rule cannot see an option after an operand. The guards therefore fail closed: every `rm`, `git rm`, `git worktree remove`, and `gio trash` prompts, and only macOS `trash` is allowed (owner decision). A `prompt` goes to `approvals_reviewer = "auto_review"`, a reviewer subagent, not always to a person. `docks.rules` carries the destructive-command guards (`docks-kit docs command-guards`); `cli/test/unit/codexSyncRules.test.ts` parses every one-line `prefix_rule`, alternatives lists and `justification` included, and asserts each guard's decision. Keep each rule on one line so that parser sees it.
75
75
  - `codexSync.ts syncRules, kit rules manifest` records the sha256 of the kit bytes of each rules file that the kit deploys in `~/.codex/.kit-managed-rules`, in `<sha256> <file name>` lines. The manifest sits outside `rules/`, so Codex never reads it as a rule file. Sync writes each rules file through `atomicWrite.ts replaceFile`, so an interrupted write keeps the previous file. Codex loads only `*.rules` files, so it never loads the `*.tmp` sibling. Before the deploy, sync writes the manifest ahead. It keeps the recorded set of hashes of each name and adds the new kit hash of each shipped name. So a name whose kit bytes change has two lines. A sync that stops after a deploy thus leaves a record of the bytes that it wrote. A stale kit rules file keeps its old rules active. When the SoT drops a rules file, every sync therefore removes the deployed copy while its hash is in the recorded set of its name. A dropped file that the user edited stays, keeps its set, and gets a warning on each sync. Every recorded hash is a hash of kit bytes, so no record lets cleanup remove bytes that the user wrote. Cleanup reads and removes only recorded names that the SoT no longer ships. It never reads or removes a file that the SoT does not ship and no sync recorded, such as `default.rules`. Sync still deploys each name that the SoT ships over an existing file, recorded or not. It backs up a file that differs to `<name>.bak` and replaces it. After the deploy and the cleanup, the final manifest holds one line per shipped name, plus the set of each kept dropped name. Sync replaces the manifest through a temporary file and a rename, so an interrupted sync keeps the previous complete manifest. `--dry-run` prints `[dry-run] remove <file>` and writes nothing.
@@ -89,13 +89,13 @@ omp SoT notes:
89
89
  - `ompFileDeploy.ts syncMergedYaml` deep-merges `config.yml` through `ompYaml.ts mergeOmpConfig` and `models.yml` through `mergeOmpModels`. Both wrap one generic mapping merge; only the config wrapper prunes stale `retry.fallbackChains` wildcards and merges `bash.patterns`.
90
90
  - `SoT/.omp/config.yml` `bash.patterns` holds the destructive-command guards. omp's `CRITICAL_BASH_PATTERNS` check returns a bare safety override before a matching prompt rule, and omp's default `tools.approvalMode: yolo` ignores a bare override, so a prompt on a critical command becomes an allow, in subagents too (`omp://approval-mode.md`). Every `rm` shape that check can flag is therefore a deny: recursive `rm`, `sudo` anywhere before `rm` (also by path), any `rm` with an `r` or `f` option before an absolute path, and `--no-preserve-root`. By owner decision (0.20.9), the recursive-`rm` deny also covers any prefix (a `* rm ` copy of each bare cluster rule, so `env X=1 rm -rf b`, `xargs rm -rf`, `find . -exec rm -rf {} +`, and `git rm -r dir` are denied) and the `-ri`, `-ir`, `-Ri`, and `-iR` clusters for bare `rm`, `*/rm`, and `* rm`. By owner decision (0.21.0), 17 more denies cover bulk removal that can empty the home directory: `find -delete` (bare, `*/find`, `* find`), `-exec`, `-execdir`, `-ok`, and `-okdir` with `rm` or `*/rm`, and `xargs rm` in six shapes. No glob can express "inside the temporary directory", so `docks-kit scratch remove` checks that path in code. A `$` or a backtick anywhere after `git push` in the same command is a deny, not a prompt (0.22.5, #120). A variable can expand to a `+` refspec or `--force`, and the matcher reads the text before the shell expands it. The agent reads the refusal and writes the ref literally, so no person is asked. The two rules are `*git push *$*` and ``*git push *`*``. Their leading `*` keeps a shell keyword such as `do`, a wrapper such as `env`, or a path before `git` from hiding the push. By owner decision (0.22.8, #126), the kit ships no omp prompt rule: `bash.patterns` holds only the denies plus the `trash *` allow, so every command that no deny matches runs without a question, in subagents too. Since 0.20.9 this run-without-question behavior covers a non-denied `rm`, `git rm`, `git worktree remove`, and `gio trash`. Since 0.21.0 it also covers `docks-kit scratch remove`. Since 0.22.8 it also covers `git clean`, `git reset --hard`, a hook skip, and every `git push` that no deny matches. The force-push denies match only `git push --force` or `-f` directly before `origin main` or `origin master`, so another spelling, such as `git push origin main --force`, runs. A prompt rule that the user adds survives sync and still asks. The kit plugin trigger rules in `SoT/.omp/plugin/rules/` carry the agent rules for these commands. `git-safety.md` tells the agent to write each `git push` ref literally and to push the current branch with `git push -u origin HEAD`. Claude Code `permissions.ask` and Codex `docks.rules` keep their ask tiers. The list order is deny, then `trash *`. omp takes the first match. `ompYaml.ts mergeBashPatterns, grouped rule merge` keeps user rules and groups all rules as kit deny, user deny, user prompt, kit prompt, other user rules, kit allow, so a user deny or prompt precedes every kit allow. It treats a deployed rule equal to a kit rule (approval case-folded, `match` whitespace collapsed) as the kit copy, so a re-sync is a no-op. A rule the kit stops shipping stays as a user rule, so retire one through `ompRemovals.ts OMP_RETIRED_BASH_RULES`. `cli/test/lib/ompBashPatterns.ts` ports omp 18.17.0's matcher, critical patterns, and yolo resolution; `cli/test/unit/ompBashPatterns.test.ts` asserts each command's final outcome and the merge. omp subagents reject every `prompt`, and the `eval` tool stays unrestricted by owner decision.
91
91
  - `ompSync.ts` runs `ompRemovals.ts syncOmpRemovals, retired-key inventory` right after the config merge, and that pass force-prunes retired kit-owned keys from `~/.omp/agent/config.yml` on every sync, without `--reconcile`. The pass is required because `mergeOmpConfig` is additive, so removing a key from the SoT alone never removes it from a deployed file. A key retired with a recorded value is pruned only while the deployed value still matches that value, so a user edit survives; a key retired outright, such as `providers.webSearchOrder`, is pruned at any value. The same pass removes each `bash.patterns` rule listed in `OMP_RETIRED_BASH_RULES`: the 39 file-removal rules retired in 0.20.9, the four `git push` prompts for a `$` or backtick word that 0.22.5 replaced with denies, and the 492 prompt rules retired in 0.22.8 (`RETIRED_PROMPT_MATCHES`). It compares with the merge's kit-rule normalization (approval case-folded, `match` whitespace collapsed), so a user rule with a different `match` or `approval` survives.
92
- - `cycleOrder` ends with `astra` as its fifth stop. `modelRoles.astra` is `openai-codex/gpt-6-astra:xhigh`, and `modelTags.astra` is visible. Astra and Fable fall back to each other through concrete selectors. `modelRoles.fable` is `anthropic/claude-fable-5-2:medium`, with visible `modelTags.fable`. The hidden `switch_fable` role uses the same Fable selector and keeps an empty fallback chain.
93
- - `modelRoles.task` is `anthropic/claude-opus-5-5:high`, and `retry.fallbackChains.task` falls back to `openai-codex/gpt-6.1-sol:high`. The kit relies on omp's default `task.enableEffort: false`: a caller `effort` outranks the role's level suffix, and `hi` maps to the top of the model ladder, so with the field on, `hi` spawns ran at `task.maxEffort` (`xhigh`) instead of the role level. With the field off, every spawn runs its role level, and `task.maxEffort` bounds only a per-call `model` with a `:level` suffix. By owner decision (0.22.22), the omp advisor is removed: the SoT has no `advisor` block, `advisor` role, or `advisor` fallback chain. The plan-lifecycle verification loop replaces a second model on every turn. A session that runs `/advisor on` takes its advisor model from the `slow` role. `smol`, `commit`, and `tiny` use `openai-codex/gpt-6-luna`. The kit-defined hidden `reviewer` role is `openai-codex/gpt-6.1-terra:high`. Its retry fallback is `anthropic/claude-opus-5-5:high`. All four reviewer entries in `task.agentModelOverrides` point to `@reviewer`. `task.agentServiceTierOverrides` sets the same four agents to `inherit`, so they take the live tier of the main session for each provider. The user set the effort `high` and chose to follow the fast toggle (#134); the review benchmark of #130 chose the model. `ompOverlay.ts ROLE_ORDER, FALLBACK_ORDER` list `reviewer`, so a free session also replaces it. omp bundles the `reviewer` and `security-reviewer` agents, and the `plan-lifecycle` plugin ships `code-reviewer` and `plan-reviewer`. The `SoT/toolchain.json` omp floor is 18.4.4, the oldest release the kit tested with `gpt-6.1-sol`. That release lists and serves the model through the `openai-codex` provider. `cli/docs/omp-models.md` records the checks.
94
- - The six Opus 5.5 roles (`default`, `slow`, `plan`, `designer`, `vision`, `task`) and the `smol`, `tiny`, `commit`, and `reviewer` retry chains use `anthropic/claude-opus-5-5` at the levels the role map records. `cli/docs/omp-models.md` carries the Artificial Analysis capture behind the role map, read 2026-09-29 at Intelligence Index v4.3.2 and Coding Agent Index v1.5 from the AA comparison-page metric tables. The `reviewer`, `fable`, and `switch_fable` rows are the exceptions: they come from v5.0.1 captures read 2026-11-05 (GPT-6.1 Terra) and 2026-11-12 (Fable 5.2), and they compare only with other v5.0.1 figures. Opus 5.5 max has the highest v4.3.2 index in that topic, at 58. AA has measured speed and latency for Opus 5.5 max and every GPT-6.1 Sol level; GPT-6 Luna medium is the only unmeasured speed and latency row.
92
+ - `cycleOrder` ends with `astra` as its fifth stop. `modelRoles.astra` is `openai-codex/gpt-6.2-astra:xhigh`, and `modelTags.astra` (`GPT-6.2 Astra`) is visible. Astra and Fable fall back to each other through concrete selectors: `retry.fallbackChains.astra` holds `anthropic/claude-fable-5-2:medium`, and `retry.fallbackChains.fable` holds `openai-codex/gpt-6.2-astra:xhigh`. `modelRoles.fable` is `anthropic/claude-fable-5-2:medium`, with visible `modelTags.fable`. The hidden `switch_fable` role uses the same Fable selector and keeps an empty fallback chain.
93
+ - `modelRoles.task` is `anthropic/claude-opus-5-5:high`, and `retry.fallbackChains.task` falls back to `openai-codex/gpt-6.2-sol:high`. The kit relies on omp's default `task.enableEffort: false`: a caller `effort` outranks the role's level suffix, and `hi` maps to the top of the model ladder, so with the field on, `hi` spawns ran at `task.maxEffort` (`xhigh`) instead of the role level. With the field off, every spawn runs its role level, and `task.maxEffort` bounds only a per-call `model` with a `:level` suffix. By owner decision (0.22.22), the omp advisor is removed: the SoT has no `advisor` block, `advisor` role, or `advisor` fallback chain. The plan-lifecycle verification loop replaces a second model on every turn. A session that runs `/advisor on` takes its advisor model from the `slow` role. `smol`, `commit`, and `tiny` use `openai-codex/gpt-6-luna`. The kit-defined hidden `reviewer` role is `openai-codex/gpt-6.1-terra:high`. Its retry fallback is `anthropic/claude-opus-5-5:high`. All four reviewer entries in `task.agentModelOverrides` point to `@reviewer`. `task.agentServiceTierOverrides` sets the same four agents to `inherit`, so they take the live tier of the main session for each provider. The user set the effort `high` and chose to follow the fast toggle (#134); the review benchmark of #130 chose the model. `ompOverlay.ts ROLE_ORDER, FALLBACK_ORDER` list `reviewer`, so a free session also replaces it. omp bundles the `reviewer` and `security-reviewer` agents, and the `plan-lifecycle` plugin ships `code-reviewer` and `plan-reviewer`. The `SoT/toolchain.json` omp floor is 18.22.2, the oldest release the kit tested with `gpt-6.2-sol` and `gpt-6.2-astra`. That release lists and serves both models through the `openai-codex` provider. The floor is doctor visibility only. Sync does not check the omp catalog, so an older omp can fuzzy-match a missing id to another model. `cli/docs/omp-models.md` records the checks.
94
+ - The six Opus 5.5 roles (`default`, `slow`, `plan`, `designer`, `vision`, `task`) and the `smol`, `tiny`, `commit`, and `reviewer` retry chains use `anthropic/claude-opus-5-5` at the levels the role map records. `cli/docs/omp-models.md` carries the Artificial Analysis capture behind the role map, read 2026-09-29 at Intelligence Index v4.3.2 and Coding Agent Index v1.5 from the AA comparison-page metric tables. The `astra`, `reviewer`, `fable`, and `switch_fable` rows are the exceptions: they come from v5.0.1 captures read 2026-11-18 (GPT-6.2 Astra), 2026-11-05 (GPT-6.1 Terra), and 2026-11-12 (Fable 5.2), and they compare only with other v5.0.1 figures. Opus 5.5 max has the highest v4.3.2 index in that topic, at 58. AA has measured speed and latency for every Opus 5.5 level and every GPT-6.2 Sol and GPT-6.2 Astra row in that topic. GPT-6 Luna medium is the only unmeasured speed and latency row.
95
95
  - `modelRoles.web` is `web/firecrawl` and `retry.fallbackChains.web` carries the explicit 20-entry provider order. The kit declares both keys because the legacy `providers.webSearchOrder` key is retired: omp expands it in memory into these two keys and then drops it, and never writes that expansion back to disk. An explicit chain replaces omp's built-in web order wholesale, so every entry left out is a provider omp never tries. The owner removed the seven entries that named older models (Gemini 2.5 Flash, Claude Haiku 4.5, GPT-5.6, GPT-5.5, and Grok 4.5); keep every other provider.
96
- - `SoT/.omp/models.yml` declares Astra's full `low, medium, high, xhigh, max` ladder with `defaultLevel: xhigh` as the worked provider ladder-override example. The temporary `anthropic.modelOverrides.claude-opus-5-5` block is gone, because the shared catalog now publishes the same limits, ladder, and prices. The removal switches omp from the block's `thinking.mode: effort` (a `budget_tokens` value per level from `ANTHROPIC_THINKING`, no effort value) to the catalog's `anthropic-adaptive` (adaptive thinking with the role's effort level); every kit Opus selector names its level, so the block's bare-selector `defaultLevel: high` is not needed. `ompRemovals.ts syncOmpModelRemovals, retired-block inventory` prunes it from `~/.omp/agent/models.yml` after the models merge, but only while the deployed block still equals the shipped one. `ompYaml.ts mergeOmpModels` preserves deployed-only keys in `~/.omp/agent/models.yml`, because a user file may carry provider credentials. Whole-file replacement is wrong.
96
+ - `SoT/.omp/models.yml` declares Astra's full `low, medium, high, xhigh, max` ladder with `defaultLevel: xhigh` as the worked provider ladder-override example. The override is keyed `gpt-6.2-astra`, the model id of the `astra` role. `ompRemovals.ts syncOmpModelRemovals` prunes the old `gpt-6-astra` block from `~/.omp/agent/models.yml` while it equals the shipped block. The temporary `anthropic.modelOverrides.claude-opus-5-5` block is gone, because the shared catalog now publishes the same limits, ladder, and prices. The removal switches omp from the block's `thinking.mode: effort` (a `budget_tokens` value per level from `ANTHROPIC_THINKING`, no effort value) to the catalog's `anthropic-adaptive` (adaptive thinking with the role's effort level); every kit Opus selector names its level, so the block's bare-selector `defaultLevel: high` is not needed. `ompRemovals.ts syncOmpModelRemovals, retired-block inventory` prunes it from `~/.omp/agent/models.yml` after the models merge, but only while the deployed block still equals the shipped one. `ompYaml.ts mergeOmpModels` preserves deployed-only keys in `~/.omp/agent/models.yml`, because a user file may carry provider credentials. Whole-file replacement is wrong.
97
97
  - `SoT/.omp/AGENTS.md` carries the rule `Please remove all mannered prose.` Anthropic's Fable 5.1 prompting guide documents mannered prose as a Fable 5.1 behavior and gives that sentence as its short-version fix: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1
98
- - `cli/docs/omp-models.md` (topic `omp-models`) records the role map rationale and the Artificial Analysis snapshot behind it. Model choices change with published benchmarks, so update that topic in the same commit as a role change.
98
+ - `cli/docs/omp-models.md` (topic `omp-models`) records the role map rationale and the Artificial Analysis snapshot behind it. Model choices change with published benchmarks, so update that topic in the same commit as a role change. For each model that an omp role or retry chain uses, the topic keeps AA rows at `high` and `xhigh` at least, with the index version and the capture date.
99
99
  - `SoT/.omp/config.yml` carries only keys whose value differs from the omp default. When a kit value equals the omp default, remove the key from the SoT. Add it with its shipped value to `ompRemovals.ts OMP_RETIRED_VALUES`, because `ompYaml.ts mergeOmpConfig` is additive and only that pass prunes the deployed copy. For example, the kit ships no `compaction.thresholdTokens`, and sync prunes a deployed `-1` and the former `231200` pin. `cli/docs/omp-context.md` (topic `omp-context`) carries the reserve-based derivation and the measured evidence, so update that topic in the same commit as any compaction-setting change.
100
100
  - Sync materializes the kit omp plugin `docks-kit-omp` from `SoT/.omp/plugin/` to `~/.docks-kit/omp-plugin` (`ompKitPlugin.ts syncKitPlugin`). It sets directories to 0700, files to 0600, and the `package.json` `version` to the kit version, and it removes stale files. It keeps no backups, because the directory is kit-owned. `ompPlugins.ts syncPlugins, kit plugin link` then runs `omp plugin link <dir>`. It skips the link when the `omp plugin list --json` npm row `docks-kit-omp` already has a `path` that resolves to that directory. `--skip-plugin-refresh` does not skip the link. The `hygiene.ts` extension registers `scratch_new`, `scratch_remove`, and `tidy_report`. It adds a reminder after each bash command that created a resource, and it checks at session stop that those resources are gone. At a main-session start, it claims `~/.cache/docks-kit/tidy.stamp.claim` and runs `docks-kit tidy --apply` in the background at most once per 12 hours. Every extension subprocess runs without a shell. The `pluginUpgrades.ts` extension reads the omp plugin upgrade log. In a tool result of each main session, it adds one reminder for the logged upgrades that the session did not get before. It names only the upgrades that finished after its omp process started. The reminder tells the agent to ask the user to run `/reload-plugins`. It then tells the agent to read each plugin skill again through `skill://` before it runs a skill script. It asks for a restart instead when an upgraded plugin ships hooks, extension modules, or non-MCP tools. omp imports extensions only at startup. An omp process that started before a sync deployed `pluginUpgrades.ts` did not load it, so its sessions get no reminder. The plugin also ships the `workspace-hygiene` skill, the `/cleanup` command, and six omp Time Traveling Stream Rules (TTSR, trigger rules) in `rules/`: `file-removal.md`, `git-safety.md`, `temp-and-background.md`, `worktree-location.md`, `literal-paths.md`, and `remote-cleanup.md`. omp's `omp-plugins` provider loads `rules/*.md` from the linked plugin root (`omp://plugin-manager-installer-plumbing.md`). Each rule has a regex `condition` scoped to `tool:bash, tool:eval` that aims at the wrong form; eval is in scope because eval code can run a shell command through `sh()`. A condition reads text, not a parsed command, so it has known false positives: a quoted command word (`echo "rm x"`, or a commit message that starts with `rm`) and eval code after the closing quote of a `sh()` string. `worktree-location.md` fires on each `git worktree add`, also at the correct path, because a condition cannot read `$XDG_DATA_HOME` or the home directory. omp aborts the matching tool call before it runs, injects the rule, and retries (`omp://ttsr-injection-lifecycle.md`). Each rule has `repeatMode: always`: it fires at each later wrong form, but not on the retry that its own reminder starts. The rules reach subagents and cost no prompt tokens until they fire. They teach the correct form; the `bash.patterns` denies still enforce. omp tests each condition on the partial argument buffer after each streamed delta, so a condition must not match a prefix of a correct command: it needs a delimiter after each token it checks, never the end anchor `$`. The buffer is the argument JSON text, the intent field `i` can come before or after `command`, and a provider can put JSON white space around the colon (`{"command": "ls"}`). So each condition starts its match inside the `command` or `code` string with the lookbehind `(?<="(?:command|code)"\s*:\s*"(?:\\.|[^"\\])*)`, each scan stops at an unescaped `"` and at a newline (`\n`), a tab (`\t` in the buffer) counts as shell white space, and a scan that looks for an option stops at a ` -- ` word. `cli/test/unit/ompTriggerRules.test.ts` pins every condition through bash and through Python and JavaScript eval `sh()` calls, in the compact and the spaced JSON layouts, with some rows in the pretty layout and trigger words in the intent before and after the command. It checks that each condition has a wrong row that fires it and that each rule has `repeatMode: always`. `cli/docs/command-guards.md` (topic `command-guards`) lists the rules and the false positives. The plugin, the deployed rules, and the sweep call `docks-kit scratch` and `docks-kit tidy`. So a checkout sync must not run before the global `docks-kit` carries those commands.
101
101
  - `task.agentIdleTtlMs` is `120000`, so omp parks an idle subagent after 2 minutes instead of 7. `browser.idleCloseSec` is `600`.
@@ -33,8 +33,8 @@ constant.
33
33
 
34
34
  | Lane | Window | Reserve | Threshold | Percent of window |
35
35
  |---|---:|---:|---:|---:|
36
- | `openai-codex/gpt-6-astra`, `openai-codex/gpt-6-luna` | 272,000 | 40,800 | 231,200 | 85.0% |
37
- | `openai-codex/gpt-6.1-sol`, `openai-codex/gpt-6.1-terra` | 500,000 | 75,000 | 425,000 | 85.0% |
36
+ | `openai-codex/gpt-6.2-astra`, `openai-codex/gpt-6-luna` | 272,000 | 40,800 | 231,200 | 85.0% |
37
+ | `openai-codex/gpt-6.2-sol`, `openai-codex/gpt-6.1-terra` | 500,000 | 75,000 | 425,000 | 85.0% |
38
38
  | `anthropic/*` | 1,000,000 | 150,000 | 850,000 | 85.0% |
39
39
 
40
40
  Every lane compacts at the same fraction of its own window. The absolute
@@ -93,8 +93,8 @@ to `contextWindow - 1`. On any model whose window is at or below the pinned
93
93
  value, the reserve collapses to one token, so omp compacts only once the
94
94
  context is already full. omp 18.8.8 and later fall back to the reserve-based
95
95
  budget for such a pin, so this collapse does not occur there. The kit floor is
96
- omp 18.4.4, so a supported host can still have the clamp. The reserve-based
97
- default always leaves at least 15% of the window free.
96
+ omp 18.22.2, later than 18.8.8, so a supported host no longer has the clamp.
97
+ The reserve-based default always leaves at least 15% of the window free.
98
98
 
99
99
  No current role model has a window below 272,000, so the hazard was latent
100
100
  rather than live. Not pinning removes it entirely, including for any future
@@ -112,7 +112,7 @@ key turns idle compaction off on fresh installs.
112
112
 
113
113
  | Setting | Value | Reason |
114
114
  |---|---|---|
115
- | `extendedContext` | `false` | omp 17.4.0 added the setting with the default on, and a later release changed the default to off. The setting controls the window omp reports for a model with a long-context limit. On omp 18.10.4, `omp models openai-codex` lists `openai-codex/gpt-6.1-sol` at 500,000 tokens with the setting off and 922,000 tokens with it on. The Codex lane uses this model for the `default`, `task`, and `vision` fallbacks. The `reviewer` role uses `openai-codex/gpt-6.1-terra`, which omp 18.16.0 also lists at 500,000 tokens with the setting off and which has the same 272,000-token long-prompt rule. OpenAI bills the whole request at the long-prompt rate when the input is above 272,000 tokens. For GPT-6.1 Sol, that rate is input $4.00 instead of $2.00, cache read $0.20 instead of $0.10, cache write $5.00 instead of $2.50, and output $15.00 instead of $10.00. The cost is a cliff rather than a marginal rate. Keeping the setting off keeps the smaller window, so fewer requests cross that cliff. |
115
+ | `extendedContext` | `false` | omp 17.4.0 added the setting with the default on, and a later release changed the default to off. The setting controls the window omp reports for a model with a long-context limit. On omp 18.10.4, `omp models openai-codex` listed `openai-codex/gpt-6.1-sol` at 500,000 tokens with the setting off and 922,000 tokens with it on. The fallback chains now use `openai-codex/gpt-6.2-sol` for the `default`, `task`, and `vision` fallbacks. omp 18.22.2 lists that model at 500,000 tokens with the setting off. The kit has not checked the window that omp 18.22.2 reports for it with the setting on. The `reviewer` role uses `openai-codex/gpt-6.1-terra`, which omp 18.16.0 also lists at 500,000 tokens with the setting off and which has the same 272,000-token long-prompt rule. OpenAI bills the whole request at the long-prompt rate when the input is above 272,000 tokens. For GPT-6.2 Sol, OpenAI applies 2x the input and cache rates and 1.5x the output rate to the full request. From the base prices, that gives input $4.00 instead of $2.00, cache read $0.20 instead of $0.10, cache write $5.00 instead of $2.50, and output $15.00 instead of $10.00. The cost is a cliff rather than a marginal rate. Keeping the setting off keeps the smaller window, so fewer requests cross that cliff. |
116
116
  | `compaction.idleThresholdTokens` | `200000` | omp default. |
117
117
  | `compaction.idleTimeoutSeconds` | `300` | omp default. |
118
118
  | `compaction.methodOrder` | `["remote","snapcompact","handoff","shake","soft"]` | omp default, remote first. Since omp 18.1.18 the Anthropic server-side lane handles Anthropic models, so snapcompact is a fallback rather than the primary path. |
@@ -18,7 +18,7 @@ against.
18
18
  | `tiny` | `openai-codex/gpt-6-luna` | low | 21 | $0.0045 | 2.18 s |
19
19
  | `fable` | `anthropic/claude-fable-5-2` | medium | 56 (v5.0.1) | $3.80 (v5.0.1) | 11.23 s (v5.0.1) |
20
20
  | `switch_fable` | `anthropic/claude-fable-5-2` | medium | 56 (v5.0.1) | $3.80 (v5.0.1) | 11.23 s (v5.0.1) |
21
- | `astra` | `openai-codex/gpt-6-astra` | xhigh | 52 | $2.31 | 126.90 s |
21
+ | `astra` | `openai-codex/gpt-6.2-astra` | xhigh | 62 (v5.0.1) | $2.52 (v5.0.1) | 52.89 s (v5.0.1) |
22
22
  | `reviewer` | `openai-codex/gpt-6.1-terra` | high | 42 (v5.0.1) | $0.26 (v5.0.1) | 10.91 s (v5.0.1) |
23
23
  | `web` | `web/firecrawl` | n/a | n/a | n/a | n/a |
24
24
 
@@ -26,19 +26,38 @@ The table reports the measured Artificial Analysis figures for each assigned
26
26
  model and level. It states no motive that the config or omp's own
27
27
  documentation does not establish. AA has not measured output speed or latency
28
28
  for GPT-6 Luna medium, so that row carries `n/a` for TTFT. AA measures no web
29
- search provider, so the `web` row carries no figures. The `reviewer`, `fable`,
30
- and `switch_fable` rows come from later captures at Intelligence Index v5.0.1.
31
- The 2026-09-29 capture predates GPT-6.1 Terra and Fable 5.2. Compare those
32
- rows only with other v5.0.1 figures (see GPT-6.1 Terra and Claude Fable 5.2
33
- below). A review benchmark chose the model of the `reviewer` role
34
- (DocksDocks/public#130), not these figures. The tier of that role follows the
35
- main session (DocksDocks/public#134).
29
+ search provider, so the `web` row carries no figures. The `astra`, `reviewer`,
30
+ `fable`, and `switch_fable` rows come from later captures at Intelligence
31
+ Index v5.0.1: GPT-6.1 Terra from 2026-11-05, Fable 5.2 from 2026-11-12, and
32
+ GPT-6.2 Astra from 2026-11-18. The 2026-09-29 capture predates these models.
33
+ Compare those rows only with other v5.0.1 figures (see GPT-6.2 Astra, GPT-6.1
34
+ Terra, and Claude Fable 5.2 below). A review benchmark chose the model of the
35
+ `reviewer` role (DocksDocks/public#130), not these figures. The tier of that
36
+ role follows the main session (DocksDocks/public#134).
36
37
 
37
38
  The role map carries no Coding Agent Index column. That index measures specific
38
39
  agent harnesses and model levels, not omp roles. The snapshot section below
39
- lists the relevant entries, including Codex with GPT-6.1 Sol high.
40
+ lists the relevant entries, including Codex with GPT-6.2 Sol and GPT-6.2 Astra
41
+ at high and xhigh.
40
42
 
41
- ### GPT-6 Sol and GPT-6 Luna availability
43
+ ### AA rows for each role model
44
+
45
+ For each model that an omp role or retry chain uses, this topic records AA
46
+ rows at `high` and `xhigh` at least. Each set of rows names its Intelligence
47
+ Index version and its capture date. A role or chain change adds the missing
48
+ rows in the same commit. On 2026-11-18, the role map and the retry chains use
49
+ six models, and each has both rows:
50
+
51
+ | Model | Index version | Capture date |
52
+ |---|---|---|
53
+ | Claude Opus 5.5 | v4.3.2 | 2026-09-29 |
54
+ | GPT-6 Luna | v4.3.2 | 2026-09-29 |
55
+ | Claude Fable 5.2 | v5.0.1 | 2026-11-12 |
56
+ | GPT-6.1 Terra | v5.0.1 | 2026-11-05, speed re-measured 2026-11-18 |
57
+ | GPT-6.2 Sol | v5.0.1 | 2026-11-18 |
58
+ | GPT-6.2 Astra | v5.0.1 | 2026-11-18 |
59
+
60
+ ### OpenAI model availability in omp
42
61
 
43
62
  The `openai-codex` selectors for GPT-6 Sol and GPT-6 Luna were deployed on
44
63
  2026-09-22, while the rollout of both models was still in progress. The owner
@@ -74,6 +93,15 @@ and the request returned HTTP 401 `refresh_token_invalidated`. The Codex
74
93
  model cache was last fetched on 2026-09-22 and does not list `gpt-6.1-sol`,
75
94
  so that Codex run did not establish model availability.
76
95
 
96
+ On 2026-11-18, omp 18.22.2 listed `gpt-6.2-sol` with a 500K context window
97
+ and `gpt-6.2-astra` with a 272K context window. Both had low, medium, high,
98
+ xhigh, and max. An `omp -p --mode json` run on each selector at `:low`
99
+ recorded that same id as the serving model. omp 18.22.2 listed no GPT-6.2
100
+ Terra and no GPT-6.2 Luna. To recheck, run `omp models openai-codex`; the
101
+ `gpt-6.2-sol` and `gpt-6.2-astra` rows must appear. Sync deploys the selectors
102
+ without a catalog check. On an omp release without these ids, omp can resolve
103
+ them by fuzzy match to another model without a warning, as on 2026-09-22.
104
+
77
105
  What omp's settings catalog establishes about these roles:
78
106
 
79
107
  - `cycleOrder` lists the roles the model switcher cycles. `fable` is the
@@ -94,7 +122,7 @@ provider. "OpenAI fast mode" below describes how that tier changes. The bundled
94
122
  `reviewer` and `security-reviewer` agents ship with omp, and the
95
123
  `plan-lifecycle` plugin ships `code-reviewer` and `plan-reviewer`.
96
124
  `retry.fallbackChains.reviewer` holds `anthropic/claude-opus-5-5:high`, and
97
- `retry.fallbackChains.task` holds `openai-codex/gpt-6.1-sol:high`; both are
125
+ `retry.fallbackChains.task` holds `openai-codex/gpt-6.2-sol:high`; both are
98
126
  cross-vendor fallbacks.
99
127
 
100
128
  The review benchmark of DocksDocks/public#130 ran on 2026-11-05. It used 24
@@ -117,10 +145,10 @@ recall difference down to -0.07 (bound -0.05). Both intervals include 0, and
117
145
  the user judged the difference as noise.
118
146
 
119
147
  `retry.fallbackChains.astra` holds `anthropic/claude-fable-5-2:medium`.
120
- `retry.fallbackChains.fable` holds `openai-codex/gpt-6-astra:xhigh`.
148
+ `retry.fallbackChains.fable` holds `openai-codex/gpt-6.2-astra:xhigh`.
121
149
  Each deliberate cycle stop falls to the other vendor. Without these explicit
122
150
  chains, `retry.fallbackChains.default` would send either stop to
123
- `openai-codex/gpt-6.1-sol:high`.
151
+ `openai-codex/gpt-6.2-sol:high`.
124
152
  Chain entries are concrete selectors, not role aliases, so this pair cannot
125
153
  recurse. The hidden `switch_fable` chain stays empty.
126
154
 
@@ -175,12 +203,13 @@ comes from the GPT-5.6 Sol high column of the GPT-6.1 Sol high comparison.
175
203
  Index composition changes between versions; figures from an earlier capture
176
204
  cannot be mixed with these.
177
205
 
178
- The Coding Agent Index page (`/agents/coding-agents`) lists 30 harness, model,
179
- and level entries. It covers GPT-6.1 Sol with Codex at low, medium, high,
180
- xhigh, and max; Opus 5.5 and Fable 5.1 with Claude Code at max; and Astra and
181
- Luna with Codex at max. It also covers Grok Build with Grok 4.7 at xhigh,
182
- Antigravity SDK with Gemini 3.8 Flash at high, and Opencode with GLM-5.3 and
183
- Kimi Code CLI with Kimi K3 without levels. The relevant published scores are:
206
+ At v1.5 on 2026-09-29, the Coding Agent Index page (`/agents/coding-agents`)
207
+ listed 30 harness, model, and level entries. It covered GPT-6.1 Sol with Codex
208
+ at low, medium, high, xhigh, and max; Opus 5.5 and Fable 5.1 with Claude Code
209
+ at max; and Astra and Luna with Codex at max. It also covered Grok Build with
210
+ Grok 4.7 at xhigh, Antigravity SDK with Gemini 3.8 Flash at high, and Opencode
211
+ with GLM-5.3 and Kimi Code CLI with Kimi K3 without levels. The relevant
212
+ published v1.5 scores are:
184
213
 
185
214
  | Harness and model | Coding Agent Index |
186
215
  |---|---:|
@@ -196,7 +225,8 @@ Kimi Code CLI with Kimi K3 without levels. The relevant published scores are:
196
225
  | Codex - GPT-6 Luna (max) | 41 |
197
226
 
198
227
  These are coding-agent scores for specific harnesses and levels, not scores
199
- for the omp role map.
228
+ for the omp role map. The v1.5 table stays as history. Coding Agent Index v2.1
229
+ below has the 2026-11-18 entries for GPT-6.2 Sol and GPT-6.2 Astra.
200
230
 
201
231
  Column meanings:
202
232
 
@@ -212,6 +242,29 @@ Column meanings:
212
242
  - **TB 4.0** - Terminal-Bench 4.0, one of the index evaluations (ten at v4.3.2,
213
243
  nine at v5.0.1).
214
244
 
245
+ ### Coding Agent Index v2.1
246
+
247
+ Source: `https://artificialanalysis.ai/agents/coding-agents`, read on
248
+ 2026-11-18 at Coding Agent Index v2.1. This version gives equal weight to
249
+ FrontierCode-AA 2.0, Terminal-Bench 4.1, and SWE-Atlas-QnA-AA, and lists 39
250
+ entries. The page draws the scores only in charts. The values below are the
251
+ `indexScore` fields of the page data, on a 0 to 1 scale. The cost and agent
252
+ wall time columns are the mean per task from the same data. This capture did
253
+ not observe the rendered integer values. Do not compare these values with the
254
+ v1.5 integers above.
255
+
256
+ | Harness and model | indexScore | Cost/task | Agent wall time s |
257
+ |---|---:|---:|---:|
258
+ | Codex - GPT-6.2 Astra (max) | 0.5347 | $5.117 | 1506.7 |
259
+ | Codex - GPT-6.2 Astra (xhigh) | 0.5312 | $3.198 | 992.4 |
260
+ | Codex - GPT-6.2 Astra (high) | 0.5228 | $2.372 | 947.4 |
261
+ | Codex - GPT-6.2 Sol (max) | 0.4862 | $1.216 | 1051.4 |
262
+ | Codex - GPT-6.2 Sol (xhigh) | 0.4979 | $0.898 | 755.8 |
263
+ | Codex - GPT-6.2 Sol (high) | 0.4746 | $0.764 | 635.4 |
264
+
265
+ GPT-6.2 Sol scores higher at `xhigh` than at `max`. v2.1 lists GPT-6.1 Sol
266
+ and GPT-6.1 Terra with Codex only at max.
267
+
215
268
  ### Claude Opus 5.5 (Anthropic) - `anthropic/claude-opus-5-5`
216
269
 
217
270
  | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
@@ -294,7 +347,74 @@ baseline at v4.3.2.
294
347
  Price: $10.00 in, $50.00 out, $0.25 cache hit per 1M. Context 1M. All levels
295
348
  run with fallback.
296
349
 
297
- ### GPT-6 Astra (OpenAI) - `openai-codex/gpt-6-astra`
350
+ ### GPT-6.2 Sol (OpenAI) - `openai-codex/gpt-6.2-sol`
351
+
352
+ Source: `https://artificialanalysis.ai`, read on 2026-11-18 at Intelligence
353
+ Index v5.0.1. The rows come from the rendered metric table of
354
+ `/models/comparisons/gpt-6-2-sol-vs-gpt-6-1-terra`, which shows every level of
355
+ both models. Tokens per task and index tokens come from the data of the model
356
+ page `/models/gpt-6-2-sol`, because the comparison table does not show them.
357
+ The other values in that data match the rendered table. These rows compare
358
+ only with other v5.0.1 figures, not with the v4.3.2 rows.
359
+
360
+ | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
361
+ |---|---:|---:|---:|---:|---:|---:|---:|
362
+ | max | 57 | $1.01 | 40k | 61M | 85 | 66.70 | 60% |
363
+ | xhigh | 56 | $0.65 | 24k | 33M | 81 | 26.63 | 56% |
364
+ | high | 55 | $0.56 | 20k | 23M | 78 | 9.63 | 59% |
365
+
366
+ AA and OpenAI list $2.00 input, $0.10 cached input, $2.50 cache write, and
367
+ $10.00 output per 1M tokens. OpenAI also lists a 1,050,000-token context,
368
+ 922,000 maximum input tokens, and 128,000 maximum output tokens
369
+ (<https://developers.openai.com/api/docs/models/gpt-6.2-sol>). Prompts above
370
+ 272K input tokens cost 2x input and cache rates and 1.5x output rates for the
371
+ full request. The knowledge cutoff is 2026-06-01. OpenAI released the model on
372
+ 2026-11-17 (<https://developers.openai.com/api/docs/changelog>). Codex uses a
373
+ 500,000-token default context (<https://developers.openai.com/codex/models>).
374
+ On 2026-11-18, omp 18.22.2 listed `gpt-6.2-sol` with the same 500K context
375
+ window, 128K maximum output, and the levels low, medium, high, xhigh, and max.
376
+
377
+ `retry.fallbackChains.default` and `.task` moved from GPT-6.1 Sol `high` to
378
+ GPT-6.2 Sol `high`, and `.vision` moved at `medium` (DocksDocks/public#175).
379
+ On 2026-11-18, the v5.0.1 data of the model page `/models/gpt-6-1-sol` listed
380
+ GPT-6.1 Sol high at index 50 and $0.59 per task. GPT-6.2 Sol high has 55 and
381
+ $0.56.
382
+
383
+ ### GPT-6.2 Astra (OpenAI) - `openai-codex/gpt-6.2-astra`
384
+
385
+ Source: `https://artificialanalysis.ai`, read on 2026-11-18 at Intelligence
386
+ Index v5.0.1. The rows come from the rendered metric table of
387
+ `/models/comparisons/gpt-6-2-astra-vs-gpt-6-2-sol`. Tokens per task and index
388
+ tokens come from the data of the model page `/models/gpt-6-2-astra`. These
389
+ rows compare only with other v5.0.1 figures, not with the v4.3.2 rows.
390
+
391
+ | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
392
+ |---|---:|---:|---:|---:|---:|---:|---:|
393
+ | max | 62 | $4.00 | 41k | 66M | 53 | 170.99 | 64% |
394
+ | xhigh | 62 | $2.52 | 22k | 27M | 54 | 52.89 | 64% |
395
+ | high | 60 | $2.07 | 18k | 20M | 56 | 22.07 | 60% |
396
+
397
+ The `max` and `xhigh` rows both round to 62. The page data gives 62.14 and
398
+ 61.56.
399
+
400
+ AA and OpenAI list $8.00 input, $0.40 cached input, $10.00 cache write, and
401
+ $40.00 output per 1M tokens. OpenAI also lists a 1,050,000-token context,
402
+ 922,000 maximum input tokens, and 128,000 maximum output tokens
403
+ (<https://developers.openai.com/api/docs/models/gpt-6.2-astra>). The
404
+ long-prompt rule of GPT-6.2 Sol also applies. The knowledge cutoff is
405
+ 2026-05-18. OpenAI released the model on 2026-11-17, in the same changelog
406
+ entry as GPT-6.2 Sol. Codex uses a 272,000-token default context. On
407
+ 2026-11-18, omp 18.22.2 listed `gpt-6.2-astra` with the same 272K context
408
+ window, 128K maximum output, and the levels low, medium, high, xhigh, and max.
409
+
410
+ `modelRoles.astra` and `retry.fallbackChains.fable` moved from GPT-6 Astra
411
+ `xhigh` to GPT-6.2 Astra `xhigh` (DocksDocks/public#175). GPT-6 Astra below
412
+ has the v5.0.1 figures for a like-for-like comparison.
413
+
414
+ ### GPT-6 Astra (OpenAI) - previous generation
415
+
416
+ The role map and the retry chains no longer use GPT-6 Astra. This table stays
417
+ as the v4.3.2 baseline for the previous `astra` role.
298
418
 
299
419
  | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
300
420
  |---|---:|---:|---:|---:|---:|---:|---:|
@@ -307,7 +427,17 @@ run with fallback.
307
427
  Price: $10.00 in, $50.00 out, $1.00 cache hit per 1M. Context 1M. Knowledge
308
428
  cutoff 2026-04-30. AA publishes no non-reasoning Astra row.
309
429
 
310
- ### GPT-6.1 Sol (OpenAI) - `openai-codex/gpt-6.1-sol`
430
+ On 2026-11-18, the v5.0.1 data of the AA model page `/models/gpt-6-astra`
431
+ listed GPT-6 Astra `xhigh` at index 55 and $3.66 per task, with 44 t/s and an
432
+ 83.77 s TTFT. It listed `high` at index 55, $3.24, 44 t/s, and 40.21 s. At
433
+ `xhigh`, GPT-6.2 Astra scores 62 against 55 and costs $2.52 against $3.66 per
434
+ task. It produces 54 against 44 t/s, with a TTFT of 52.89 s against 83.77 s.
435
+ AA marks GPT-6 Astra deprecated, so its comparison pages return HTTP 404.
436
+
437
+ ### GPT-6.1 Sol (OpenAI) - previous generation
438
+
439
+ The role map and the retry chains no longer use GPT-6.1 Sol. This table stays
440
+ as the v4.3.2 baseline for the former `task` role and retry fallbacks.
311
441
 
312
442
  | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
313
443
  |---|---:|---:|---:|---:|---:|---:|---:|
@@ -323,7 +453,9 @@ OpenAI also lists $2.50 per 1M cache-write tokens, a 1,050,000-token context,
323
453
  272K input tokens cost 2x input and cache rates and 1.5x output rates for the
324
454
  full request. The knowledge cutoff is 2026-04-30. The API effort ladder is
325
455
  `low, medium, high, xhigh, max`, with `medium` as the default. There is no
326
- `none`, `minimal`, or non-reasoning level.
456
+ `none`, `minimal`, or non-reasoning level. On 2026-11-18, omp 18.22.2 still
457
+ served `gpt-6.1-sol`, and the OpenAI deprecations page did not list it. AA
458
+ marks it deprecated.
327
459
 
328
460
  ### GPT-6.1 Terra (OpenAI) - `openai-codex/gpt-6.1-terra`
329
461
 
@@ -346,6 +478,14 @@ For comparison at the same index version, AA lists GPT-6.1 Sol high at index
346
478
  TTFT, and 52% on Terminal-Bench 4.0. The same Sol high row read 50, $0.32, and
347
479
  57.26 s at v4.3.2, which shows why the two captures do not mix.
348
480
 
481
+ On 2026-11-18, the release comparison page
482
+ `/models/comparisons/gpt-6-2-sol-vs-gpt-6-1-terra` showed new speed and TTFT
483
+ figures for GPT-6.1 Terra at v5.0.1. `high` ran at 65 t/s with a 10.50 s
484
+ TTFT, `xhigh` at 65 t/s with 30.33 s, and `max` at 69 t/s with 146.49 s. The
485
+ index, cost, and token figures were unchanged. AA measures speed and latency
486
+ continuously, so the table keeps the 2026-11-05 rows. The speed comparison
487
+ below uses the 2026-11-18 figures.
488
+
349
489
  AA and OpenAI list $1.00 input, $5.00 output, and $0.05 cached input per 1M
350
490
  tokens. OpenAI also lists a 1,050,000-token context, 922,000 maximum input
351
491
  tokens, and 128,000 maximum output tokens
@@ -353,6 +493,30 @@ tokens, and 128,000 maximum output tokens
353
493
  Fast mode costs 2x Standard. The knowledge cutoff is 2026-06-01. The API effort
354
494
  ladder is `none, low, medium, high, xhigh, max`, with `medium` as the default.
355
495
 
496
+ ### Output speed: GPT-6.2 Sol and GPT-6.1 Terra
497
+
498
+ GPT-6.2 Sol has the higher output speed at `high` and at `xhigh`.
499
+ The figures come from one page,
500
+ `/models/comparisons/gpt-6-2-sol-vs-gpt-6-1-terra`, read on 2026-11-18 at
501
+ Intelligence Index v5.0.1.
502
+
503
+ | Metric | GPT-6.2 Sol | GPT-6.1 Terra |
504
+ |---|---:|---:|
505
+ | `high` output speed | 78 t/s | 65 t/s |
506
+ | `xhigh` output speed | 81 t/s | 65 t/s |
507
+ | `high` TTFT | 9.63 s | 10.50 s |
508
+ | `xhigh` TTFT | 26.63 s | 30.33 s |
509
+ | `high` end-to-end time | 42.65 s | 43.91 s |
510
+ | `xhigh` end-to-end time | 85.03 s | 86.53 s |
511
+
512
+ At `high`, Sol produces about 19% more tokens per second than Terra. At
513
+ `xhigh`, the gain is about 24%. Sol also has the lower TTFT and end-to-end
514
+ time at both levels.
515
+
516
+ The `reviewer` role stays on GPT-6.1 Terra `high`. The #130 benchmark chose
517
+ Terra by escaped defects, not by speed. omp 18.22.2 lists no GPT-6.2 Terra. A
518
+ move of the reviewer to GPT-6.2 Sol needs a new benchmark.
519
+
356
520
  ### GPT-6 Luna (OpenAI) - `openai-codex/gpt-6-luna`
357
521
 
358
522
  | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
@@ -373,7 +537,7 @@ has measured both at the other levels.
373
537
  ### GPT-5.6 Sol (OpenAI) - previous generation
374
538
 
375
539
  The role map no longer uses GPT-5.6 Sol. This table stays as the measured
376
- baseline for the GPT-6.1 Sol `task` retry fallback.
540
+ baseline for the former GPT-6.1 Sol `task` retry fallback.
377
541
 
378
542
  | Level | Index | Cost/task | Tokens/task | Index tokens | Speed t/s | TTFT s | TB 4.0 |
379
543
  |---|---:|---:|---:|---:|---:|---:|---:|
@@ -407,12 +571,16 @@ Price: $0.20 in, $1.20 out, $0.02 cache hit per 1M. Context 1M.
407
571
 
408
572
  `task` runs `anthropic/claude-opus-5-5:high`, the same model and level as
409
573
  `default`. The owner moved it from `openai-codex/gpt-6.1-sol:high` on
410
- 2026-10-09. GPT-6.1 Sol high is now the `task` retry fallback. The owner uses
411
- Astra only for main orchestration, so Astra has a dedicated `astra` cycle stop
412
- at `xhigh`. This comparison keeps the retired Astra-low, GPT-5.6 Sol, and
413
- GPT-6.1 Sol choices beside the current Opus level.
414
-
415
- | Metric | Astra low, retired | GPT-5.6 Sol high, retired | GPT-6.1 Sol high, fallback | GPT-6.1 Sol max | Opus 5.5 high, current |
574
+ 2026-10-09. GPT-6.1 Sol high then became the `task` retry fallback. Since
575
+ DocksDocks/public#175, the fallback is GPT-6.2 Sol high
576
+ (`openai-codex/gpt-6.2-sol:high`). The owner uses Astra only for main
577
+ orchestration, so Astra has a dedicated `astra` cycle stop at `xhigh`. This
578
+ comparison keeps the retired Astra-low, GPT-5.6 Sol, and GPT-6.1 Sol choices
579
+ beside the current Opus level. It is the v4.3.2 capture of 2026-09-29, so it
580
+ does not compare with the v5.0.1 GPT-6.2 rows. GPT-6.2 Sol above has the
581
+ figures for the current fallback.
582
+
583
+ | Metric | Astra low, retired | GPT-5.6 Sol high, retired | GPT-6.1 Sol high, former fallback | GPT-6.1 Sol max | Opus 5.5 high, current |
416
584
  |---|---:|---:|---:|---:|---:|
417
585
  | Intelligence Index | 46 | 42 | 50 | 52 | 54 |
418
586
  | Cost per Index task | $0.82 | $0.81 | $0.32 | $0.72 | $1.82 |
@@ -443,9 +611,9 @@ Opus 5.5 high scores 54 against GPT-6.1 Sol high at 50, 57% against 52% on
443
611
  Terminal-Bench 4.0, and 1705 against 1471 on AA-Briefcase. It costs $1.82
444
612
  against $0.32 per index task. Its answer TTFT is 52.85 s against 57.26 s.
445
613
 
446
- The `astra` cycle stop runs xhigh: index 52, $2.31 per index task, and
447
- 126.90 s TTFT. AA publishes a Coding Agent Index entry for Astra only at
448
- max with Codex, where it scores 62.
614
+ The `astra` cycle stop runs GPT-6.2 Astra `xhigh`. At v5.0.1 it has index 62,
615
+ $2.52 per index task, and a 52.89 s TTFT (see GPT-6.2 Astra). Coding Agent
616
+ Index v2.1 lists Codex with GPT-6.2 Astra at high, xhigh, and max.
449
617
 
450
618
  ## How Astra is selected in practice
451
619
 
@@ -456,10 +624,13 @@ No subagent selects Astra automatically.
456
624
 
457
625
  `SoT/.omp/models.yml` declares the full `low, medium, high, xhigh, max` ladder
458
626
  with `defaultLevel: xhigh` under
459
- `providers.openai-codex.modelOverrides.gpt-6-astra.thinking`.
627
+ `providers.openai-codex.modelOverrides.gpt-6.2-astra.thinking`.
460
628
  The override stays as the kit's worked example of a provider ladder
461
629
  redefinition. The in-session thinking control can still select `low` for a
462
- quick answer.
630
+ quick answer. The block is keyed by model id, so it moved from `gpt-6-astra`
631
+ to `gpt-6.2-astra` with the role. `ompRemovals.ts syncOmpModelRemovals`
632
+ prunes the deployed `gpt-6-astra` block while it equals the shipped block, and
633
+ keeps a block that the user edited.
463
634
 
464
635
  - The bundled `scout` and `sonic` agents carry `model: "@smol"` and
465
636
  `thinking-level: medium` in their embedded frontmatter, so they run Luna,
@@ -493,9 +664,12 @@ The runtime per-agent measurement from fresh `omp -p` runs on 2026-09-09
493
664
  predates this change. It applies to the retired Astra-low `task` configuration,
494
665
  not the current role map.
495
666
 
496
- omp's 272k context window is the `/extended-context off` window for Astra.
497
- `/extended-context on` uses the 922k input window, 1.05M total, which matches
498
- AA's 1M.
667
+ On 2026-11-18, omp 18.22.2 listed GPT-6.2 Astra with a 272,000-token context
668
+ window. This is the `/extended-context off` window, and it equals the Codex
669
+ default context. OpenAI lists a 1,050,000-token context and 922,000 maximum
670
+ input tokens. For GPT-6 Astra, `/extended-context on` used the 922k input
671
+ window. The kit has not checked the window that `/extended-context on` gives
672
+ for GPT-6.2 Astra.
499
673
 
500
674
  ## Free session launcher (`docks-kit omp`)
501
675
 
@@ -680,10 +854,13 @@ and the ladder counts in these docs in the same commit.
680
854
  values, and exits without starting a session. `priority` is the tier omp's
681
855
  `/fast on` selects for OpenAI and OpenAI-Codex models. `--fast` over
682
856
  `ultrafast` selects `priority`. The kit ships no Ultrafast toggle: omp honors
683
- `ultrafast` for a Codex model only when its catalog `serviceTiers` lists it,
684
- and the omp 18.4.9 catalog lists none for GPT-6.1 Sol, Astra, or Luna. The SoT
685
- does not declare `tier.openai`, so the additive sync merge keeps the toggled
686
- value. New sessions read it; a running session changes through `/fast`.
857
+ `ultrafast` for a Codex model only when its catalog `serviceTiers` lists it.
858
+ The omp 18.4.9 catalog listed none for GPT-6.1 Sol, Astra, or Luna. On
859
+ 2026-11-18, `omp models openai-codex --json` on omp 18.22.2 listed no
860
+ `serviceTiers` for any `gpt-6` id, `gpt-6.2-sol` and `gpt-6.2-astra`
861
+ included. The SoT does not declare `tier.openai`, so the additive sync merge
862
+ keeps the toggled value. New sessions read it; a running session changes
863
+ through `/fast`.
687
864
 
688
865
  The four review agents inherit the live tier of the main session
689
866
  (`task.agentServiceTierOverrides: inherit`). In a new session after
@@ -702,12 +879,27 @@ on normal processing, use `docks-kit omp --fast` and a new session, not
702
879
 
703
880
  ## Maintenance
704
881
 
705
- - Refresh each per-level row from the metric table of
706
- `/models/comparisons/<level-slug>-vs-gpt-5-6-sol-high`. Check that the page
707
- title names the requested model and level. Take `max` from the bare model
708
- slug. Do not read values from the chart payloads of the model pages: each
709
- chart holds only about 20 models, so a missing value there does not mean
710
- that AA has not measured it.
882
+ - Refresh each per-level row from the rendered metric table of a release
883
+ comparison page, for example
884
+ `/models/comparisons/gpt-6-2-sol-vs-gpt-6-1-terra` and
885
+ `/models/comparisons/gpt-6-2-astra-vs-gpt-6-2-sol`. Each page shows every
886
+ level of both models. Take tokens per task and index tokens from the data
887
+ of the model page, for example `/models/gpt-6-2-sol`, and check that its
888
+ other values match the rendered table. A model-page chart holds only about
889
+ 20 models, so a value missing there does not mean that AA has not measured
890
+ it.
891
+ - On 2026-11-18, the per-level pattern
892
+ `/models/comparisons/<level-slug>-vs-gpt-5-6-sol-high` returned HTTP 404. A
893
+ comparison with a model that AA marks deprecated also returned HTTP 404, for
894
+ example `gpt-6-1-sol`, `gpt-6-astra`, or `gpt-5-6-sol`. Read such a model
895
+ from the data of its model page.
896
+ - Keep AA rows at `high` and `xhigh` at least for each model that an omp role
897
+ or retry chain uses, with the index version and the capture date (see AA
898
+ rows for each role model). Add the missing rows in the same commit as a
899
+ role or chain change.
900
+ - After a role or chain change, run `omp models openai-codex` and check that
901
+ it lists every `openai-codex` id that the roles and chains name. Sync does
902
+ not check the omp catalog.
711
903
  - Refresh the Coding Agent Index from `/agents/coding-agents`.
712
904
  - Record the index version with the numbers. AA changes index composition
713
905
  between versions, so a score from another version is not a comparison.
@@ -104,6 +104,10 @@ const OMP_RETIRED_KEYS: ReadonlyArray<string> = ["providers.webSearchOrder"];
104
104
  * served the id as an empty stub. The catalog now publishes the same limits,
105
105
  * prices, and ladder. The block's `defaultLevel: high` applied only to a bare
106
106
  * selector, and every kit selector names its level.
107
+ *
108
+ * `gpt-6-astra` carried the kit's worked ladder override while the `astra`
109
+ * role ran GPT-6 Astra. The role moved to GPT-6.2 Astra, and the override
110
+ * moved with it to the `gpt-6.2-astra` key, so no role uses this block.
107
111
  */
108
112
  const OMP_RETIRED_MODEL_BLOCKS: ReadonlyArray<readonly [string, unknown]> = [
109
113
  [
@@ -122,6 +126,16 @@ const OMP_RETIRED_MODEL_BLOCKS: ReadonlyArray<readonly [string, unknown]> = [
122
126
  },
123
127
  },
124
128
  ],
129
+ [
130
+ "providers.openai-codex.modelOverrides.gpt-6-astra",
131
+ {
132
+ thinking: {
133
+ mode: "effort",
134
+ efforts: ["low", "medium", "high", "xhigh", "max"],
135
+ defaultLevel: "xhigh",
136
+ },
137
+ },
138
+ ],
125
139
  ];
126
140
 
127
141
  /** Each two-letter short-option cluster with a first letter from `first` and a second from `second`. */