inbrace-config 0.8.0

Minor Changes

• Add the Claude Haiku 4.5 → Haiku 5.5 transition audit, started with /inbrace-config:audit-haiku-5-5, with its known traps, change digest and regression cases. (#136 by @ropdias)

Patch Changes

• Sonnet 5 → 5.5 audit: price Claude Sonnet 5.5 cache reads at USD 0.10 per million tokens, half Sonnet 5's, and flag a price entry copied from Sonnet 5. (#138 by @ropdias)

Full changelog: https://github.com/inbrace-tech/claude-skills/blob/inbrace-config@0.8.0/plugins/inbrace-config/CHANGELOG.md

inbrace-config 0.7.0

Minor Changes

• The audit reaches no network: it reads the change digest shipped with each transition, drops the drift stage and the docs-only mode, and the contribution invite only gives a link. (#114 by @ropdias)

Patch Changes

• The Sonnet eval fixture's sample clients no longer read ANTHROPIC_API_KEY themselves and leave it to the SDK; the plugin directory's validation held that read for review. (#103 by @ropdias)

• Two known traps follow the docs as they read today: Opus P04's passage is re-recorded, and Sonnet P25 now says Opus 5.5 reads Sonnet 5.5 thinking blocks on the Claude API and Google Cloud. (#112 by @ropdias)

• Every passage the knowledge files quote was re-checked against Anthropic's docs on 2026-10-06 and still holds, so each verified date now reads 2026-10-06. (#123 by @ropdias)

• Each transition now ships a change digest beside its known traps: what changed between the two models, each item with its page and a short quoted passage. No skill reads it yet. (#111 by @ropdias)

• The README and descriptions say what the audit now does: it reads the docs as recorded and dated in the plugin, and fetches and sends nothing. The eval suite fails a run that reaches the network. (#115 by @ropdias)

Full changelog: https://github.com/inbrace-tech/claude-skills/blob/inbrace-config@0.7.0/plugins/inbrace-config/CHANGELOG.md

inbrace-config 0.6.3

Patch Changes

• The plugin now ships a listing icon, which Anthropic's plugin directory reads when the plugin is first submitted. (#97 by @ropdias)

• The Sonnet eval fixture's sample help-desk client no longer reads a token from the environment, which the plugin directory's validation held for review. (#98 by @ropdias)

• The audits now flag a line in a rule, memory or command file that restates an agent's or skill's effort, such as a routing-table row, for re-test with the file it mirrors. (#95 by @ropdias)

• The plugin folder now has its own README: what the audit does, its commands, what it fetches and sends, troubleshooting, support and the privacy policy. (#99 by @ropdias)

Full changelog: https://github.com/inbrace-tech/claude-skills/blob/inbrace-config@0.6.3/plugins/inbrace-config/CHANGELOG.md

inbrace-config 0.6.2

Patch Changes

• Transition audit: the plan warns when a reasoning-exposing line sits in always-loaded memory, and a refused agent's batches are then read in the session. (#88 by @ropdias)

• Transition audit: a value the project keeps on purpose for a reason the docs bear out, such as a Sonnet 5 fallback variable, is already decided, not a change. (#88 by @ropdias)

• Transition audit: re-anchored the Opus P07 source and re-recorded the Sonnet P04 passage from today's docs. (#88 by @ropdias)

• Eval suite: the key labels the reasoning-in-the-reply rows by trap id and never quotes them, and a Sonnet fixture docstring names the loop whose conversation it escalates. (#88 by @ropdias)

• Transition audit: new known traps for the source model id in Sonnet API code (P44), and for Opus API code a max_tokens sized with thinking off (P29) and requests with no effort (P30). (#88 by @ropdias)

• Transition audit: Opus P20 also reads launch commands and CI workflows, and P23 fires for an agent already on Opus 5.5 that preloads an Opus 5 skill. (#88 by @ropdias)

• Transition audit: Opus P07 recognises rules copied from the Opus 5 guide without naming a model, and P11 fires on a multi-app agent that writes records even with a specific brief. (#88 by @ropdias)

• Transition audit: cites reasoning-exposing lines by number, never reading them, and recovers from a model refusal instead of stopping. (#88 by @ropdias)

• Transition audit: two traps on one construct each keep their own line, and Sonnet's P03 effort re-test reads the memory, rules and hooks areas where P02 fires. (#88 by @ropdias)

• Transition audit: a price in a skill's text shows its amount, not the invocation's argument, and check-norms rejects a dollar sign before a digit in a skill. (#88 by @ropdias)

• Transition audit: the verifier looks for a file with Glob before it calls one missing, and keeps a reference that could sit outside the project. (#88 by @ropdias)

Full changelog: https://github.com/inbrace-tech/claude-skills/blob/inbrace-config@0.6.2/plugins/inbrace-config/CHANGELOG.md

inbrace-config 0.6.1

Patch Changes

• Without AskUserQuestion, the audit closes as a headless run naming --scope and --mode instead of listing its questions. (#82 by @ropdias)

• Known traps now cover the computer-use agent loop, unattended API agent loops, source-model rationales and capability claims, and model notes in memory files. (#81 by @ropdias)

• The Sonnet 5.5 audit flags instructions that message an agent while it works, a risk P27 covered only in Claude API code. (#73 by @ropdias)

• Adds a local claude plugin eval suite for the Sonnet 5.5 and Opus 5.5 audits: seeded fixtures and deterministic graders from their keys. (#83 by @ropdias)

• When a project refuses the close's cleanup, the audit leaves the working files, names them and never retries the deletion another way. (#75 by @ropdias)

• The Sonnet 5.5 audit proposes removing MAX_THINKING_TOKENS=0 and a per-tool-call countdown hook instead of listing them as re-tests. (#72 by @ropdias)

• Findings no longer rest on the user's own settings, the verifier keeps documented requirements as changes, and older residue stays out of the final list. (#80 by @ropdias)

• The finding verifier checks and corrects every file:line a finding cites, in its notes as well as its anchor. (#74 by @ropdias)

Full changelog: https://github.com/inbrace-tech/claude-skills/blob/inbrace-config@0.6.1/plugins/inbrace-config/CHANGELOG.md

inbrace-config 0.6.0

Minor Changes

• The audit checks every known trap's doc passage against the page as it reads today, and reports traps whose passage changed as possibly stale, quoting the current text. (#45 by @ropdias)

• When a run learns something the audit's knowledge lacks, the close invites you to contribute it: a generic, anonymised issue draft, shown first and sent only if you choose. (#45 by @ropdias)

• One audit for every model transition, /inbrace-config:audit-model-transition <source> <target>, reads today's docs, explores freely, then checks known traps; the Opus and Sonnet commands run it. (#45 by @ropdias)

Patch Changes

• The plan's cost estimate counts every part of a run from measured runs — fixed load, docs, files, findings and verifiers — and states the cost in dollars. (#60 by @ropdias)

• The audit writes each run to .model-audits/<target>-<date>/, ignored by git, instead of the protected .claude/, so it runs without per-file prompts and headless. (#58 by @ropdias)

• Both audits now propose the exact-or-longest-prefix selector fix for prefix-delivered prompting (P05, P22) as a change you can apply; re-testing the delivered text stays a re-test. (#33 by @ropdias)

• Rewrite the plugin's norm and trap histories in the repository's one-line refs layout; no norm, trap or behaviour changes. (#45 by @ropdias)

• Add the discover stage of the coming transition-agnostic audit, which judges files against today's docs, and let batch-auditor work from its digest; the current audits are unchanged. (#45 by @ropdias)

• Add the orchestrator and plan stage of the coming transition-agnostic audit; no command starts them yet, so nothing changes for users. (#45 by @ropdias)

• Add the report and apply stages of the coming transition-agnostic audit; no command starts them yet. (#45 by @ropdias)

• Add the known-traps stage of the coming transition-agnostic audit, and let finding-verifier check each finding's doc passage and verify one shard of files. (#45 by @ropdias)

• Add the knowledge files the coming transition-agnostic audit reads, for Opus 5 → 5.5 and Sonnet 5 → 5.5: every pattern row, with the doc passages it rests on. (#45 by @ropdias)

• Known traps are checked in every area they apply to (a missed P13 in skills), and a run writes only inside its folder, reporting any file it left elsewhere. (#62 by @ropdias)

Full changelog: https://github.com/inbrace-tech/claude-skills/blob/inbrace-config@0.6.0/plugins/inbrace-config/CHANGELOG.md

inbrace-config 0.5.0

• New skill, /inbrace-config:audit-sonnet-5-5, audits a setup for the move from Claude Sonnet 5 to Sonnet 5.5 with the same stages and gates as the Opus audit; it proposes the model claude-sonnet-5-5 without writing an effort level, and the batch-auditor and finding-verifier agents now serve both audits, reading the transition from the brief.
• Both audits find the old tier's prompting delivered to the new model by a prefix match in gate and check scripts as well as hooks, and read and count only the scripts that match; they keep recommending full scope when the skills carry the old model's tuning, estimate parallel time from the slowest auditor's batch count, and recommend showing the diff or stopping when the project binds a pin to other files.
• Both audits also read templates, triggers and a worktree's gitignored instruction files, find Claude API code by its SDK imports, anchor each finding on the line that carries it with one line per pattern, compute the total cost as fixed plus marginal, and recommend by a fixed ordered list; the verifier merges only exact duplicates.
• Both audits take --scope, --mode and --stop-at-report arguments that answer the plan gate and stop at the report, so they can run headless; no argument applies changes.
• Both audits also read model-dependent code and docs — price tables, prompting-guide indexes, model-id selectors and transcript parsers — and add rows for a source-model skill preloaded into a moving agent, the Claude Code version floor, content-based fallback, price tables, model docs and text-only parsers; a prefix selector's fix names exact or longest matching, project memory counts for a missing instruction unless the agent sets omitClaudeMd, every row states a confidence, and older residue takes the id P00.

Install or update: /plugin marketplace add inbrace-tech/claude-skills, then /plugin install inbrace-config@inbrace; existing installs: claude plugin update inbrace-config@inbrace.

Full history: CHANGELOG.md