agentleFS
nabondancenabondance/strata/skills/strata-interpret/SKILL.md

strata-interpret

Skill1 starsChanged 8 days agoMITView on GitHub

Interpret Strata output about Git-tracked AI-agent configuration — an `analyze` JSON or HTML report, a `strata diff` comparison in JSON or Markdown, the SVG image, or the badge. Use when a user asks what changed, what a figure or report panel means, or what Strata can and cannot establish.

  • Reads credentials

The file

Files
SKILL.md

Interpret a Strata report

Read the user's Strata JSON report and check schemaVersion before interpreting fields. A history report from strata analyze has revisions; a comparison from strata diff has base, head, and comparison and is described near the end, with the other outputs. If the user has only the HTML report, it embeds the same history report as JSON: the first line containing const data = holds it on one line, so parse the text after const data = up to that line's final ; (later occurrences belong to embedded code). This skill describes history schema version 22; if the report has another version, inspect its fields and avoid assuming this schema still applies. Version 21 stores every revision's componentCount, estimatedTokens, skillListingTokens, skillListingUnresolved, providerTokenTotals, and changes (empty when nothing changed), and leaves out only firstParent and categoryTotals when they repeat the previous revision. Version 20 and earlier store firstParent and categoryTotals in every revision, have no symlinkTarget, keep trailing punctuation in an import's path, and store a null target without external for every path no tracked file had, even one written as a path, so they list more imports as missing. Version 19 and earlier have Codex counts only with --precise, and their measurement may be an OpenAI encoding, with a model name, instead of the estimate. Version 18 and earlier have no lineAge, and set shallow more broadly, so it can be true for a complete history. Version 17 and earlier have no mcp server lists. Version 16 and earlier have no upFront labels. Version 15 and earlier do not resolve Claude @path imports, so they have no imports, importRoot, or imported components, and their totals can be lower for the same history; version 14 and earlier have no copilot provider and do not recognize REVIEW.md; version 13 and earlier have no cursor provider. Version 12 and earlier have no skill listing or skillListingTokens; version 11 and earlier have no submodules. Version 10 and earlier have no estimate provider accuracy, credential, recipient, alsoSentTo, wrapperTokens, or fallbacks; version 9 and earlier have no readers, contextFile, or geminiContextFiles, and labeled .agents/skills files codex instead of shared. Versions 8 and earlier repeated a full components list inside every revision; version 7 and earlier have no gemini provider and no provider counts.

Use revisions in array order: every first-parent commit from the root (or the shallow boundary) to the analyzed head is one revision, so nothing is sampled, and a commit that touched no recognized file still appears, without changes and with its predecessor's totals; the number of revisions is not the number of configuration changes. Each revision has a commit id, date (the author date, which need not increase after a rebase), componentCount, estimatedTokens, and changes from its predecessor. Every field but id, date, geminiContextFiles, and lineAge is left out when it repeats the previous revision: an absent firstParent is the previous revision's id; an absent componentCount, categoryTotals, estimatedTokens, skillListingTokens, skillListingUnresolved, or providerTokenTotals equals the previous revision's, so take it from the nearest revision at or before it that has it; and absent changes mean the revision changed nothing. A revision that changed nothing therefore holds only id and date, plus geminiContextFiles while the Gemini settings name context files; that field keeps its own meaning (absent means GEMINI.md only) and is not carried forward, so its presence alone is not a change. The first and the last revision always have every field. Never read an absent field as zero or as missing data. The first revision's changes are relative to an empty baseline. A revision's totals count every file once, shared files included, so they describe overall growth without rebuilding any file list; a per-provider figure needs the replay below and the reading rules further down. If shallow is true, earlier history is unavailable. If ignore is present, it lists the --ignore patterns the user gave (.gitignore-style globs, kept verbatim even in a shareable report): every tracked path they match was left out of every revision as if it were not in the tree, so it is in no total, an import of it is missing, and the totals cover only the rest of the repository. Say so when quoting a total, and do not describe the matched folders' configuration as absent from the repository. If submodules is present, its entries are the Git submodules at the last revision, whose files were not analyzed. An entry with configFiles above zero holds that many recognized agent-configuration files, absent from every total, so do not describe the report as the repository's complete agent configuration; null means the submodule could not be checked, which depends on the local checkout. A merge appears as one revision measured against its first parent, not as a full branch history.

A revision does not list its files. The top-level components array holds every distinct file version in the history once (a path together with the blob stored there), ordered by path, so one path can appear several times with different blobs. Each change has a component field: the index in components of the version now at path, or null when the file was removed. To get the files present at a revision, start with an empty map from path to component and apply the changes of every revision from the first up to that one, in order: delete previousPath if the change has one, then delete path if component is null, otherwise set path to components[component]. The result must have componentCount entries; if it does not, the replay is wrong. Never read components as the inventory of any single revision.

A component's provider is the tool whose path convention it matches (claude, codex, gemini, cursor, copilot), shared for paths more than one tool reads, or unknown for files under .agents/ outside a skill directory. Strata recognizes only these tools' paths; another tool's files, such as Cline's .clinerules, are absent unless a Claude import reaches them, so their absence says nothing about whether the repository has them. A skill component has a skill root and a manifest or resource role. Group a revision's rebuilt files by skill.root to describe a skill directory: its manifest's estimated tokens, and its resources' count and bytes. This is another view of files already counted: do not add it to categoryTotals or count resource bytes as tokens. A directory without a manifest is a set of tracked resources, not evidence that a runnable skill exists. A manifest's skill.listing measures the text an agent keeps in context before the skill is used, under the report's measurement: the frontmatter name and description for Codex, Gemini CLI, Cursor, and Copilot; for Claude Code, name (by default the directory name) and description (by default the first nonempty body line) with when_to_use appended, the two cut together at 1,536 characters. Its status is listed with its tokens, hidden (0 tokens; disable-model-invocation: true in a .claude/skills or .cursor/skills manifest), or unresolved (tokens: null: Strata could not read the frontmatter, or a tool that requires name and description did not find both, so do not assume a cost). A manifest keeps the listing of the tool that owns its folder (.agents/skills follows Codex and Gemini CLI, which ignore disable-model-invocation); Cursor's and Copilot's own listings of .claude/skills and .agents/skills are not measured and can differ, so their up-front figures use the owner's listing. A revision's skillListingTokens sums the listings and skillListingUnresolved counts the unresolved manifests it leaves out; report that count beside the total. Listings mostly repeat text already in estimatedTokens, so never add the two; a Claude skill without name is listed by its directory name, so a listing can exceed its manifest, and a long when_to_use can make a listing much larger than its description. They exclude the agent's own framing and any shortening to fit a listing budget. Use them to tell the always-listed cost of skills apart from their full stored size, not as proof of what a session loaded.

A component with upFront: "listed" is a root instruction file (CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md, AGENTS.md, AGENTS.override.md, GEMINI.md, .github/copilot-instructions.md) or a root rule whose frontmatter loads it at launch (a .claude/rules file without paths, a .cursor/rules file with alwaysApply: true); upFront: "unresolved" marks one that may load at launch but that Strata could not decide, such as a rule with unreadable frontmatter or the legacy .cursorrules. To describe what a provider lists up front at a revision, for a session started at the repository root, take the replayed files that provider reads and keep those labeled listed (a rule only for the tool that owns it), drop AGENTS.md for Claude when a root Claude file is present and for Codex when a non-empty AGENTS.override.md is (it is skipped, neither up front nor on demand), add for Claude the files reached by imports from the root CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md, and any AGENTS.md it keeps (chains do not start at rules, though they continue through a rule an import reaches), even files another tool owns, then add the listed tokens of every skill manifest it reads, including skill folders in subdirectories such as tools/bot/.gemini/skills; only <skills folder>/<name>/SKILL.md is a manifest, so a SKILL.md grouped one level deeper is a resource with no listing. Everything else it reads is on demand: nested instructions, path-scoped rules, skill manifests, commands, agents, and memory. Files labeled unresolved are counted on demand; report how many there are beside the figures. Both figures are stored text, not loaded context or billing, and do not claim settings outside the repository, sessions started in a subdirectory, or command and agent descriptions were resolved.

A component with provider shared is read by more than one tool; its readers lists which (AGENTS.md: Claude and Codex; .agents/skills: Codex and Gemini CLI). A single SKILL.md placed directly in .agents/skills is Gemini CLI's only, labeled gemini. In a report where some revision tracks Cursor files, cursor is also in the readers of AGENTS.md, .agents/skills, and .claude/skills (which stay labeled by their own tool), but Cursor reads them only at revisions that hold a component with provider cursor; do not attribute them to Cursor at other revisions. A report without Cursor files says nothing about whether the team uses Cursor. Provider copilot (GitHub Copilot) works the same way, separately: in a report with Copilot files, copilot is also in the readers of AGENTS.md, .agents/skills, .claude/skills, and of CLAUDE.md, REVIEW.md, .claude/agents, .claude/rules, .mcp.json, and GEMINI.md, which keep their own provider; Copilot reads them only at revisions that hold a component with provider copilot. A copilot label means some Copilot client (GitHub.com chat, cloud agent, code review, CLI, or an IDE) is documented as reading the path; do not say which one did. REVIEW.md is code review guidance only (Claude Code's Code Review and Copilot code review). A Copilot .instructions.md file applies only where its applyTo matches, which Strata does not check. A component with legacy: true (the root .cursorrules, or a Copilot .chatmode.md) is a format its provider still reads but documents as superseded. A Cursor rule's stored size says nothing about whether it applied: rules can be always applied, matched by glob, requested by the agent, or manual. Components with contextFile (agents-md for AGENTS.md, gemini-md for GEMINI.md) are read by Gemini CLI only as the revision's geminiContextFiles lists them; a revision without that field means GEMINI.md only, and an empty list means the settings name no recognized context file, so Gemini reads neither. Do not attribute an unlisted context file to Gemini at that revision. Gemini's reading of AGENTS.md is carried by contextFile, not readers, so gemini is absent from its readers. Copilot reads GEMINI.md whatever Gemini's settings say. Claude Code's reading of AGENTS.md may also depend on conditions Strata does not check. Shared files are counted once in estimatedTokens.

Claude Code loads the files a CLAUDE.md or CLAUDE.local.md (and an AGENTS.md it reads) imports with @path, and their imports in turn, up to four hops. A component with imported: true is in a revision's inventory only because such an import reaches it there: Markdown is provider claude, category instructions, and counted in the totals; anything else is source-only other. When an import is added or dropped, its file appears in changes as added or removed although the file may not have changed, so do not describe that as the file being created or deleted. On an import root (any CLAUDE.md, CLAUDE.local.md, or AGENTS.md, at any depth) or any Markdown file an import reaches, a rule or skill manifest included, imports lists the imports in that file version, one per distinct in-repository target at its first line plus every external one, with the repository path it names as target; a file no import reaches is not parsed, so a rule's own @ lines count only once something imports the rule; target is null with external: true when the import points outside the repository, and otherwise the path as written, resolved. When no tracked file or symlink ever had that path, the target is kept only if the import is written as a path (a / or an extension such as .md, .json, or .yaml): an npm scope such as @mastra/server reads as an import, never resolves, and is not a broken file reference. A bare word no file has, such as @alice, has target: null without external; it is likely prose, and Strata does not store its text. Strata drops trailing punctuation (.,:;!?, quotes, and a closing bracket the path does not open) from the path. symlinkTarget, when present, is the file that target leads to as a Git symlink somewhere in the history (up to four links); the import reaches that file at a revision where target itself has no file. importRoot marks an import root that has imports. Whether an import resolves depends on the revision: replay the inventory, start from components with importRoot and no contextFile (any CLAUDE.md or CLAUDE.local.md; REVIEW.md is not one), follow imports to targets (or, failing those, symlinkTargets) present in the inventory by the shortest chain, then start again from importRoot components with contextFile: "agents-md" those chains did not reach. An import from a file four hops out is past the depth limit, whatever its target; otherwise one whose target is absent, or null without external, is missing. The report names these statuses resolved, missing, external, and depth-limit. Files reached only from AGENTS.md are loaded only if Claude Code reads that AGENTS.md, which Strata does not check. Report unresolved imports separately and do not count a missing target's size. An import is a link between stored files, not proof that a session loaded the target.

Explain growth or reduction using the relevant revisions, paths, categories, and providers. estimatedTokens sums instructional-file counts. Check measurement.family, accuracy, model, encoding, and implementation before describing them. In version 22, measurement is always the utf8-bytes-divided-by-four-v1 comparison estimate, for every provider, and estimatedTokens uses it. providerMeasurements describes separate counts for Codex (on every run with Codex files: OpenAI's o200k_base encoding, counted locally, whose model names that encoding) and, with --precise, Claude and Gemini on components (providerTokenCounts) and revisions (providerTokenTotals); Cursor and Copilot are never counted separately. Each has an accuracy (api-counted, encoding-exact, or estimate), credential names where a remote count's credential came from (such as ANTHROPIC_API_KEY, never the secret), and fallbacks explains why a more precise method was not used. api-counted is the provider endpoint's own count, which Anthropic documents as an estimate; call it an API count, not exact. recipient and alsoSentTo name the hosts that received instruction text. A --precise API count excludes the endpoint's message wrapper, wrapperTokens per file, so it measures the stored text alone; an api-counted measurement without wrapperTokens (the removed --provider-api, or schema 10 and earlier) includes it unless every file was blank, and encoding-exact or estimate counts have no wrapper. A revision's providerTokenTotals sums every file that provider reads at that revision, without the up-front skipping, so Claude's includes an AGENTS.md beside a root CLAUDE.md. Never add counts of different providers together. bytes measures source size. A component with estimatedTokens: null is source-only and contributes no tokens; sourceOnlyCount counts these files. Do not turn settings, hooks, MCP files, or skill resources into prompt-token estimates.

A component with mcp is an MCP file (.mcp.json, .cursor/mcp.json, .github/mcp.json, .vscode/mcp.json) or a root .gemini/settings.json or a .codex/config.toml that mentions MCP servers. status: "listed" gives the servers that file version defines, each with a name, a transport (local: the tool starts a process on the developer's machine; remote: it connects to a URL; or unknown), and disabled: true when the file turns it off; an empty list means none. status: "unresolved" means Strata could not read the servers, either because the file does not parse or because it uses syntax Strata's stricter reader does not accept (such as inline Codex mcp_servers tables): say its servers are unknown, not absent. To describe a revision's servers, replay its inventory and read mcp on the files present, by provider; when no file present has mcp, the tracked configuration mentions no MCP server. A defined server is not a loaded one: some tools ask before starting a project server and others start it once the folder is trusted, approvals can sit in settings or outside Git, and user configuration is not in the report. disabled covers Codex enabled = false and Gemini CLI's mcp.excluded and mcp.allowed, not approvals kept elsewhere. The report never measures the tool definitions a server sends, and never stores commands, URLs, or environment values, so do not estimate a server's context cost or guess what it connects to from its name.

A revision's lineAge records what it did to the age of the instruction text: added, the stored bytes (UTF-8 for text) of trimmed, nonblank lines first present at that revision, and removed, bytes it removed grouped by from, the index of the revision they had appeared at. It covers every component with stored text (estimatedTokens not null: instructions, rules, skills, agents, commands, memory, imported Markdown) across the whole recognized inventory, so it cannot be split by provider or path. Lines are matched by exact text across all of those files, so a line moved to another file keeps its age, a reworded line is a removal and an addition, and a line that disappears and returns is new again, including the text of a file that leaves the inventory because an import was dropped and returns when it is restored. Each copy of a repeated line is tracked; when copies are removed the newest go first. To get the text at a revision by origin, start with an empty map and, for every revision in order up to it that has lineAge, subtract each removed entry from its from and set the revision's own index to added. The bytes at a target whose origin is at or before a baseline revision are the baseline's text still present, so their share of the baseline's total is its survival. The report's own headline uses the first revision with instruction lines as the baseline and the last revision as the target. lineAgeMethod names this method (exact-trimmed-line-multiset-stored-bytes-v1); no line text or hash is stored. If shallow is true, the history starts where a shallow fetch cut it, so text attributed to the first revision may be older than it. These are bytes, not tokens; never add them to estimatedTokens. Survival says text stayed in the repository, not that an agent read or followed it, and exact matching misses rewording and can match a short line such as --- by coincidence.

Use changes to explain what each revision did: kind is added, removed, modified, or renamed. renamed appears only when Git pairs two recognized paths, and imported files are never renamed. deltaTokens describes the change in the stored-text estimate, not whether a change improved quality. A rename, a move into skills, or a claimed refactor may need content evidence beyond the report. Do not infer that a skill was loaded, a rule was active, or an agent used a component in a particular session.

If the user asks about something they saw in the HTML report, answer from the same data; every panel is derived from the JSON. The HTML report and the terminal summary number revisions from 1, so their revision N is revisions[N - 1], and a JSON index such as lineAge from: i is their revision i + 1. The tabs are All files, one <Provider> + shared tab per provider that owns a file or is among a file's readers (its own files, the shared files it reads, and, for Cursor and Copilot, other tools' files at revisions where that tool is configured), Unknown when there are unknown files, and Shared only when there are shared files. A tab therefore does not show that the repository configures that tool: an AGENTS.md alone gives a Codex tab. Tab totals overlap, so never add them. In a tab, a file can appear or disappear without changing, when Gemini settings change or the first Cursor or Copilot file arrives or leaves. With provider counts (Codex always, Claude and Gemini with --precise), a Claude, Gemini, or Codex tab shows that provider's providerTokenCounts (unless that provider fell back to the byte estimate) in the masthead, chart, treemap, map, Skill bundles Full text, Compare revisions, and file table, and Count with switches back to the estimate. Panels by heading:

  • The masthead gives the selected revision, its tokens, its tracked files, and Last change, the view's token change at the latest revision at or before it that changed a tracked file (the revision Previous change compares), with that revision's date beneath it.
  • Stored instruction history charts estimatedTokens by category, at most 100 bars, each the last revision of its group, starting at the first revision with a file in the view; a source-only category's legend entry shows a file count with a hollow swatch.
  • Configuration at this revision is a treemap of the view's files by stored-text tokens; source-only and zero-token files get no area and are counted below it, a solid outline marks a file added at this revision and a dashed one an edited or renamed file.
  • Configuration map has no field of its own: it lays the replayed files out by folder, with a bar for each file's share of the view's stored-text tokens, tags rows new, edited, or renamed, counts unresolved imports on the importing row, and draws a line only for each resolved import whose two files are in the view, the arrow at the imported file. A closed folder takes its files' lines, dashed; an import between two files of one closed folder is counted under the map instead. It shows a folder containing a file and a file importing another, not which files a session loaded; in a non-Claude tab the lines are still Claude imports. A shareable report aliases folder and file names but keeps the tree, so its map has the same folders.
  • Instruction age replays lineAge for all files, whatever the tab (its count then ends "· all files"), by week, month, quarter, or year of author date, choosing the finest that gives at most eight periods, or one per year when even years give more; periods exist only where text was added, so there are no empty ones; "half added by" is the byte-weighted median origin, and the share "still there" is survival since the first revision with instructions.
  • At session start is the up-front and on-demand split above, always under the report's measurement and from all files; an agent tab shows only that agent with its Read at start list. Instructions is the up-front files' tokens, Skill summaries the listings of the skill manifests it reads, Total at start the sum, When needed the on-demand tokens, and "Maybe read at start, counted as when needed" the unresolved files.
  • Skill bundles groups by skill.root. The HTML calls a listing a Summary; Full text is the manifest's tokens, Extra files and bytes its resources, and Missing under Main file a directory without a manifest.
  • Claude imports lists every import from all files with its status, unresolved ones first; the status "outside the repository" is external and "too deep" is depth-limit; the Imports column shows the path the import names, including a missing one written as a path, and, for an import that reached a file through a symlink, the link's path, an arrow, and the file; an external import shows "—", or "outside the repository" past the depth limit; a missing import whose target the report does not store (a bare word such as @alice, or in schema version 20 and earlier any path no file had) shows "not a tracked path". Read external on that import in the component's imports to tell it from an outside one. Depth is the hop count.
  • MCP servers reads mcp on the view's files.
  • Compare revisions compares a baseline (by default Previous change, the revision before the latest change at or before the target; Previous revision is the literal one before) with the selected revision, showing only groups that changed, under By provider, By category, and Changed components. It matches files by provider, category, and path, so a rename is a removal and an addition even though changes detects renames; Δ summary is the listing change, and "Instruction text kept" is survival from the baseline to the target, from all files (marked "(all files)" in a provider tab), shown only when the baseline is earlier than the target.
  • Tracked components lists the view's files; filters do not change totals, per-provider columns appear with provider counts, and "—" means that provider does not read the file at this revision or has no count for it, as for a source-only file.

The masthead's metadata line ends with ignoring N patterns when ignore is present; its tooltip lists them. Notices under the masthead flag a shallow clone, submodules that were not analyzed, and a shareable export, and Stored text, not session context explains how tokens were counted.

The default local report omits source text but includes repository paths, commit and blob IDs, MCP server names, and measurements. A privacy.mode of shareable means those identifiers were replaced with aliases; dates and measurements remain visible. Paths are aliased per segment: each folder name becomes d1, d2, … and each file stem f1, f2, … with its extension kept, while tool convention names such as .claude, skills, AGENTS.md, and SKILL.md stay, so folder depth, nesting, and which files are nested AGENTS.md are real. One alias is the same name everywhere in that report, but aliases are not cross-report identities; do not guess the names behind them. Treat all remaining metadata as potentially sensitive. If the user asks for instruction-level meaning or why wording changed, inspect the relevant Git content with their authorization instead of inventing it from the JSON.

A strata diff JSON comparison (schema version 14; version 13 and earlier do not read imports past trailing punctuation or through symlinks, so they can compare fewer imported files, 12 and earlier have Codex counts only with --precise, 11 and earlier have no limits, 10 and earlier no mcp, 9 and earlier no top-level upFront, 8 and earlier no imported files, 7 and earlier no copilot, 6 and earlier no cursor, 5 and earlier no skillListing) compares two commits' tracked trees, given as full commit IDs in base and head. An ignore list, as in a history report, names the patterns whose paths both trees leave out. It reads no history between them: there are no revisions, renames, instruction age, provider views, submodules, shallow, or geminiContextFiles, and it has no shareable mode, so paths and commit IDs are real. A change in which context files Gemini reads shows only as an edited, source-only .gemini/settings.json, a change in Gemini's upFront, and, with --precise, in Gemini's providerTotals. comparison holds the stored-text totals at both sides (beforeTokens, afterTokens, deltaTokens, file counts, and percentChange, null when the base has no tokens), byProvider and byCategory (keyed by each component's own provider, so shared stays shared), and components, only the files whose blob differs, each added, removed, or edited; a moved file is a removal and an addition. Files are matched by provider, category, and path. An imported component is compared only because a Claude import reaches it, so adding or dropping the import shows it as added or removed although the file may not have changed. comparison.skillListing gives beforeTokens, afterTokens, deltaTokens, beforeManifests, afterManifests, and the beforeUnresolved and afterUnresolved manifests those tokens leave out, and a changed manifest carries listing before and after; never add them to the stored-text totals. comparison.mcp, present when either side has a component with mcp, counts servers at both sides (disabled ones included), counts the beforeUnresolved and afterUnresolved files whose servers are not counted, and lists the servers added and removed, each by name, path, and provider; a server is identified by its file and name, so moving it to another file is a removal and an addition, and turning it off or changing its transport is not a change. The top-level upFront has one entry per provider that reads stored text at either side: before and after give tokens (up front: fileTokens plus listingTokens), onDemandTokens, fileCount, sourceOnly, skipped and skippedTokens, and the unresolved and listingUnresolved counts, for a session started at the repository root, with deltaTokens and deltaOnDemandTokens; it always uses the comparison's measurement. When present (Codex on every comparison with Codex files; Claude and Gemini only with --precise), providerMeasurements describes the counts as above, and providerTotals gives before, after, and delta tokens for each counted provider, each summing every file that provider reads, shared files included; do not compare it with the byProvider row of the same name, which holds only files labeled that provider. limits, present only when the run was given a limit such as --limit-tokens or --limit-up-front-tokens, has one entry per check: name (the flag without dashes), agent (null for the stored-text total, otherwise the agent whose session-start tokens were checked), limit in tokens, before and after, and result: over (strictly greater), within, no-agent, or replaced (a session-start limit without an agent that applied to no agent: no-agent when no agent keeps text at session start, replaced when each that does has its own limit; its agent, before, and after are null). A limit is a number of tokens the user or their workflow chose, and limits use the comparison's measurement, never providerTotals; say a figure is over the limit set for it, not that the change failed or regressed. A run with a limit passed exits with status 3 unless --on-limit warn was given; strata analyze checks --limit-tokens and --limit-up-front-tokens at the last revision but records them in no report field. The same limits apply: these are stored text, not loaded context or billing, and a growing total is not a regression by itself.

strata diff writes Markdown by default, and the GitHub Action puts it in the job summary; it does not post pull request comments. The Markdown is drawn from the same comparison but drops detail (when limits were set, a table of every check and its result comes first and matches limits): commit IDs are cut to 12 characters, a comparison with no changed file prints "No instruction files changed." in place of the category and file tables, the category table shows only changed categories and there is no byProvider table (a table of per-provider counts appears: Codex always, Claude and Gemini with --precise), only the 25 files with the largest token change are listed, largest first, so the files left out changed least, often source-only files; the session-start table shows only agents whose up-front or on-demand tokens changed (a change in unresolved or skipped counts alone reads as unchanged) and merges unresolved and listingUnresolved into one "Maybe read at start" count, and MCP names stop at ten per list. It prints no percentChange; it has no session-start line when no agent reads stored text at either commit; when no agent's tokens changed it prints only that session start is unchanged, hiding the unresolved counts, and otherwise its "Maybe read at start" line covers only the agents whose tokens changed and have a nonzero count. The MCP line appears only when either commit has a file with mcp. A "0" in the file table can be a source-only file or an edit that kept the same estimate, and an unchanged session-start figure does not rule out a change in which files an agent reads. Ask for --format json when those details matter.

The GitHub Action writes this JSON too: its json-file output, made with strata diff --json-output. On a pull request it compares GitHub's merge commit with that commit's first parent, so head is the merge commit, not the pull request's last commit, base is the base branch's tip, and the delta is what merging would change. If the job log warned that no merge commit was found or fetched, base and head are the base branch's tip and the pull request's head, and changes the base branch made after the pull request branched count, reversed, as the pull request's. The Action's limit-result output is over when an entry is over, within when every other entry checked is within, and none when no limit was checked: no-agent and replaced entries count as not checked.

Other outputs carry less. The badge from strata analyze --format badge is shields.io endpoint JSON: its schemaVersion: 1 is shields.io's, not a Strata report version, and its message is a display string of the latest revision's estimatedTokens, such as ≈12,345 tokens (with thousands separators, and ≈ because the measurement is the estimate; schema 19 and earlier could use an exact encoding, without ≈), so strip those before comparing it with the JSON; despite the "agent context" label it is stored text, and it never uses provider counts. The SVG image draws the All files history from revision totals; it names the repository and branch and shows dates, but no paths or IDs, and its header gives the latest tokens, file count, revisions, and change since the first charted revision. A --minimal SVG has no visible text or axis, so no value can be read off the drawing; only its accessible description states the latest tokens, file count, and revisions. The terminal summary from strata analyze describes the last revision with the same figures as the JSON.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.