v0.0.6

Added

• (views) scoped views — process / thread / time_range pre-filter on create_view
• (summary) hints field nudges callers toward create_view
• (views) keep_only_frames / keep_only_modules inverse-hide filter (#89)
• (views) strip_type_params preset transform (#88)
• (views) regex backreferences in rename replacements
• (views) rule hit counts on create_view and describe_view tool
• (views) stack views by composing transforms across the chain
• (views) collapse_recursion folds multi-function cycles up to length 8
• (tools) list_profiles reports view base id
• (tools) create_view MCP tool for derived profiles
• (registry) create_view derives a session from a base
• (profile) resolved_chain helper applies view transforms
• (profile) Arc-share raw tables and add transforms field
• (profile) introduce Transforms value type for views
• (query) compare_functions side-by-side asm diff (#8)
• (call_tree) surface cross-process aggregation on inverted trees
• (summary) accept process / thread / time_range filter
• (filter) warn when bare-name process= matches multiple pids
• (call_tree) surface child names in pruning markers
• (summary) add top_processes and top_threads breakdown
• (describe) cap describe_profile output and drop idle entries

Fixed

• rustfmt — missing newline before DescribeViewResult struct
• (registry) create_view caches identical views; harden tests
• (error) cap available_* lists, demangle nearest_matches, add module_not_found, route event/regex through invalid_value
• (symbolicate) count hex frames as unsymbolicated; surface per-lib outcomes
• (query) make time_range filter and summary.time_range_ms share a frame
• (query) reject empty function pattern for required/narrowing args

Other

• (release-plz) use action outputs to find release PR branch
• (changelog) record plugin bundle, cookbook, summary hints, regex backrefs
• (readme) document plugin install path alongside MCP-direct install
• extract version via cargo metadata, not grep/sed
• skill description voice, doc cleanup, manifest version guard
• --version, --help, and unknown-arg handling
• prerequisites + pollard-doctor health-check skill
• sync plugin manifest versions in release-plz PR
• marketplace.json and tag-triggered release workflow
• bundle pollard as a Claude Code plugin with skills
• surface cookbook through view_presets MCP tool (may be reverted)
• (views) cookbook of canonical hide_modules / hide_frames regex sets
• (transforms) name the keep-only placeholder inline
• cargo fmt + fix clippy::shadow_unrelated in keep_only test
• (view_stats) cargo fmt
• (matching) cargo fmt
• (view_stats) drop unneeded allow(dead_code)
• (view_stats) cargo fmt
• (spec) tighten create_view cost and lifetime wording
• (spec) document profile views
• (views) end-to-end view re-attributes hidden leaf
• (tools) clarify create_view syntax and collapse_recursion guidance
• (query) folded_stacks consumes resolved_chain
• (query) stacks_containing consumes resolved_chain
• (query) call_tree consumes resolved_chain
• (query) top_functions consumes resolved_chain
• (plan) profile views implementation plan
• Merge pull request #58 from antiguru/measure-rss
• (spec) bring design doc back in sync with shipped surface
• (query) [breaking] unify percentage field naming on {kind}_pct
• Merge pull request #77 from antiguru/claude/fix-issue-68-oPTKI
• (tools) document re: regex prefix on every function-pattern arg
• (call_tree) apply cargo fmt
• Merge pull request #54 from antiguru/issue-52-marker-no-stack
• Merge pull request #53 from antiguru/issue-51-empty-pattern

Fixed

• (symbolicate) don't enable macOS Spotlight on Linux (#56).
  SymbolManagerConfig::use_spotlight(true) was unconditional. On Linux that pushed wholesym into a macOS-shaped resolution path that ended in a dyld_shared_cache_x86_64 read, so every Linux .so failed symbolication with a stderr line like could not load symbols for "/usr/lib64/libc-2.28.so": ... /System/Library/dyld/dyld_shared_cache_x86_64 ... No such file or directory. Profiles imported from perf.data showed up degraded (top_functions, call_tree, source_for_function lost names) and the stderr volume drowned other diagnostics. Spotlight is now gated behind cfg!(target_os = "macos"), leaving the macOS dSYM-bundle lookup intact and skipping it entirely on other targets.

Changed

• breaking (query) unify percentage field naming on {kind}_pct (#65).
  summary.dominant_thread.samples_pct, summary.top_processes[].samples_pct, summary.top_threads[].samples_pct, and source_for_function[].samples_pct are renamed to self_pct so every percentage field across the surface uses the same {kind}_pct shape (self_pct / total_pct; samples reserved for cases where there's no self/total distinction at all). "Samples on this thread / process" is self time for that thread / process, and per-line source samples are self time for that line, so self_pct is the natural name and matches what top_functions already exposes. Wire-format break for callers that destructured the old name. The convention is documented next to ProcessEntry::self_pct and in the design doc.

• (tools) document the re:(?i) case-insensitive convention on every function-pattern arg (#67).
  Substring matching is case-sensitive (memcpy ≠ MEMCPY), which surprises during exploration. The regex form already supports the standard (?i) inline flag; the per-arg docs and the src/matching.rs module header now spell that out so callers don't have to discover it from the regex crate. The canonical one-liner is now Substring match by default; prefix with re:for a regex (usere:(?i)foo for case-insensitive). — same wording at every site so it stays grep-able alongside the #66 line. Default behavior is unchanged.

• (tools) document the re: regex prefix on every function-pattern arg (#66).
  The prefix that flips a function pattern from substring match to regex was discoverable through the skill description, not through per-arg docs — callers without the skill missed the feature, and callers with it had to re-find which args supported it. Every function-pattern arg now carries the same one-line "Substring match by default; prefix with re: for a regex." in its description: top_functions.filter, top_groups.filter, compare_profiles.filter, call_tree.root_function, call_tree.paths_to, folded_stacks.function_filter, stacks_containing.function, source_for_function.function, asm_for_function.function. The wording is identical at every site so it's grep-able.

Added

• (plugin) bundle pollard as a Claude Code plugin with skills (#92).
  .claude-plugin/plugin.json registers the MCP server and .claude-plugin/marketplace.json makes the bundle installable via /plugin marketplace add antiguru/pollard. Three skills ship alongside: profile-recording (samply record / samply import for perf.data / load_profile workflow with troubleshooting for empty profiles, unsymbolicated frames, and perf.data recorded without -g), view-presets (cookbook regex sets for tracing-subscriber, tokio internals, and Rust stdlib glue, plus a stacked-view recipe), and pollard-doctor (dependency-ordered diagnostic for binary-not-on-PATH, MCP-server-not-reloaded, empty-profile, and unsymbolicated-frames failure modes with exact remediation per step). Skills surface as /pollard:profile-recording, /pollard:view-presets, and /pollard:pollard-doctor and solve the discoverability gap a plain MCP tool leaves: the cookbook would otherwise require the agent to think to look it up. The plugin bundle does not ship the binary — cargo install pollard is still a prerequisite. README documents both the plugin install path and the legacy claude mcp add server-only path.

• (docs) view-presets cookbook of canonical hide_modules / hide_frames regex sets (#92).
  docs/superpowers/specs/2026-05-06-view-presets-cookbook.md collects copy-paste regex sets for the three filters users keep re-deriving by hand: tracing-subscriber walls, tokio runtime internals, Rust stdlib glue. Ships as docs (and the mirrored view-presets skill) rather than a presets= argument on create_view, so curated content can drift with upstream crates without baking yesterday's noise filters into the binary. The skill mirrors the cookbook's regex blocks; the design spec records the convention that cookbook is the source of truth and reviewers reject one-sided regex changes.

• (cli) --version, --help, and unknown-arg handling.
  pollard --version and pollard --help previously hung on stdio waiting for an MCP client. A tiny hand-rolled arg parser in src/main.rs (no new dep) now prints and exits before tokio starts; unknown args fail with exit code 2 and a hint. The MCP server path is unchanged when no args are passed. pollard-doctor's binary check uses pollard --version instead of command -v pollard so stale installs whose version doesn't match what the user expects are caught.

• (query) summary returns hints nudging callers toward create_view (#93).
  Three cheap heuristics fire on the summary response: dominant function appears recursively in stacks (suggests collapse_recursion), a single module covers >40% of stacks (suggests hide_modules), or top function names carry type parameters (suggests strip_type_params). The hints array is omitted when no heuristic fires, so a clean profile stays silent.

• (views) regex backreferences in rename replacements (#87).
  rename application now goes through a new FunctionMatcher::replace helper that delegates to regex::Regex::replace, so capture references ($1, ${name}) interpolate in the replacement string. A single rule can fold a trait-vs-inherent monomorphisation pair like re:<(.*) as .*::Schedule>::schedule => ${1}::schedule where users previously had to enumerate one rule per concrete type.

• (views) scoped views — process / thread / time_range pre-filter on create_view (#90).
  Pinning a hot pid/tid or a benchmark window once at view creation is now possible: the scope is baked into the view's stored state and applied alongside any per-call CommonFilterArgs at query time. Per-call filters must be a sub-slice of the view's scope — equality (or, for pid:, the bare-vs-suffixed strict-narrowing case) for process/thread, and [start, end] ⊆ [scope_start, scope_end] for time_range. Conflicting per-call filters are rejected with invalid_value, so a misuse surfaces as one structured error rather than a silently-ignored argument. Stacked views inherit the parent scope and are only allowed to narrow it further, by the same sub-slice rule. The composed scope round-trips through describe_view under a new scope field (thread/process/time_range, omitted when nothing is set). View ids hash the scope alongside the transforms, so two views differing only in scope register as distinct.

• (views) keep_only_frames / keep_only_modules inverse-hide transforms on create_view (#89).
  Inverse of hide_frames / hide_modules: only frames whose function name (or module name) matches at least one pattern survive, and each maximal run of non-matching frames collapses into a single <hidden> placeholder frame. The two lists OR together — a frame is kept if it matches any keep_only_* rule. Applied before hide_*, so a frame matching both a keep_only and a hide rule is dropped (hide always wins). describe_view's transforms view echoes the patterns, and rule_stats gains keep_only_frames / keep_only_modules entries whose frames_matched counts every kept frame and 0 is the same typo signal as for the other rule kinds. Composes through extend_from by appending lists (union), matching hide_*. The placeholder name is static for v1.

• (views) strip_type_params preset transform on create_view (#88).
  When strip_type_params: true, balanced <…> segments are removed from each frame's function name during transform application — OrdValBatch<RowRowLayout<((Row, Row), Ts, i64)>> collapses to OrdValBatch. The strip runs after hide_* and before rename, so user rules can target the normalized name without each rule encoding a bracket-tolerant pattern. Implemented as a hand-rolled depth counter (regex can't match balanced delimiters reliably); unmatched > outside any open bracket is preserved verbatim. describe_view's transforms view echoes the flag, and rule_stats gains a strip_type_params entry whose frames_matched counts every frame whose name actually changed — 0 is the same typo signal as for the other rule kinds. Composes through extend_from as a logical OR, matching collapse_recursion.

• (views) per-rule hit counts on create_view, plus a new describe_view tool (#85, #86).
  create_view now returns rule_stats: [{rule_index, kind, pattern, frames_matched, samples_affected}] and total_base_samples, computed by replaying every hide_frames / hide_modules / rename / collapse_recursion rule against the base's samples once at view-create time. A rule with frames_matched: 0 is the typo signal — without it, mistyped patterns failed silently and only surfaced when a downstream tool produced unchanged output. describe_view(profile_id) is symmetric with describe_profile for views: returns the immediate parent's id, the full composed transforms shape, and the same per-rule counts so callers can re-fetch them later without re-creating the view. Stats are cached on the view's ProfileSession, so describe_view is a hash-map lookup.

• (query) compare_functions(profile_id_a, function_a, function_b, profile_id_b?, ...): side-by-side asm diff with per-instruction sample counts on both sides (#8).
  Two asm_for_function calls plus an LCS pass over a normalized instruction key (registers collapsed to R, numeric immediates to IMM, mnemonic + operand shape preserved). The displayed asm text is unchanged; normalization is alignment-only — without it, register renames and differing displacements (xmm0/xmm1, rdi+rcx*8-0x38 / rdi+rcx*1-0x6000) would split nominally-equal instructions onto separate rows and the per-row sample columns would no longer line up. profile_id_b defaults to profile_id_a, so the natural "why is simd_rows_1st faster than simd_cols_1st" workflow stays a one-profile call; pass a different profile_id_b for before/after refactor diffs across two recordings. Output rows carry Option<offset|asm|samples> per side so LCS gaps (instructions only on one side, e.g. an unrolled loop with more addsds) remain visible at their original offsets. The two sides must agree on arch; a mismatch errors out rather than producing a meaningless alignment.

• (query) surface cross-process aggregation on inverted call_tree (#68).
  An inverted (inverted=true) call_tree without a process= filter pulls callers from every pid that hits a given leaf. Reasonable for "where is memcpy called from anywhere", but the same caller chain can mix time from two different processes via different code paths and the result was indistinguishable from a single-process tree. The response now carries cross_process: true and processes_in_tree: [{pid, name, samples, pct}] (biggest first; pid matches the process=pid: wire form) when the inverted aggregation crossed >1 distinct pid; single-process and process-filtered results stay silent. Top-down trees never emit the signal because their root structure already separates per-process callees.

• (query) summary accepts the standard process / thread / time_range filter (#63).
  Previously profile-wide only, so the natural follow-up after spotting a hot pid in top_processes was a hand-composed top_functions + top_modules + dominant-thread reasoning. The filter now flows through every sample-count surface in the response (total_samples, time_range_ms, duration_ms, top_processes, top_threads, top_modules, both top-functions lists, dominant_thread); recording-level fields (interval_ms, sample_rate_hz, unsymbolicated_pct, profile_start_ms) stay profile-wide because they describe the recording, not the slice. Default args reproduce the prior profile-wide behavior.

• (query) surface child names on call_tree pruning markers (#61).
  Omitted-children markers now carry top_omitted (up to 3 entries of {function, pct}, biggest first) and truncated-subtree markers carry the cutoff frame's function, so a caller can tell what was dropped and decide whether widening min_pct / max_breadth / max_depth is worth a second call.

• (query) warn when bare-name process= aggregates across multiple pids (#62).
  top_functions, call_tree, top_groups, stacks_containing, and compare_profiles now expose matched_processes: [{pid, name}] (per-side a_/b_ for compare_profiles) when a bare-name filter resolves to more than one distinct pid, listing the pids it absorbed so the caller can disambiguate via pid:N / pid:N.M. Bare-name filters that resolve to a single pid stay silent.

Fixed

• (query) align summary.time_range_ms with the time_range filter contract (#64).
  summary.time_range_ms reported absolute (boot-relative) timestamps while time_range filter args are documented as profile-relative; pasting one into the other returned no samples.
  time_range_ms is now offset by the first-sample anchor (matching the filter's reference frame), the new summary.profile_start_ms exposes that anchor so the absolute frame is one addition away, and Profile::stack_indices subtracts the anchor before gating so boot-relative samply timestamps no longer leak through.

Other

• (release) tag-triggered plugin-release.yml workflow zips .claude-plugin/, skills/, README, and LICENSE-* into pollard.skill and attaches it to the GitHub release that release-plz already cuts; a verify step rejects the tag if Cargo.toml's version doesn't match .claude-plugin/plugin.json and marketplace.json.
• (release) release-plz.yml splits into release-pr → sync → release so every release-pr run is followed by a commit that mirrors Cargo.toml's version into .claude-plugin/plugin.json and .claude-plugin/marketplace.json. release-plz force-pushes its release branch each run, so the sync commit is reapplied rather than persisting.
• (ci) new plugin-manifests job in test.yml asserts Cargo.toml version equals .claude-plugin/plugin.json and marketplace.json, catching manifest drift on regular PRs instead of waiting for the tag-time check.
• (ci) extract version via cargo metadata --no-deps --format-version 1 | jq instead of grep/sed, robust against Cargo.toml formatting changes (workspace inheritance, quoting, comments).