CHANGES

v0.0.19 (2026-10-10)

Harvey now knows the Raspberry Pi AI HAT+ 2 (hailo-ollama) as a model engine beside Ollama, learns and respects each model’s prompt-size limit, and finds models where each model system keeps them. Several behaviours change for existing setups; they come first.

Upgrade notes

Added

• Hailo (AI HAT+ 2) as an engine (harvey DR-0030). hailo-ollama is an engine label over the Ollama client, not a fourth backend, with its own hailo: section in harvey.yaml (hailo.url, default http://localhost:8000) so chat and embeddings can use different servers. The card (/dev/hailo0) and the server (GET /hailo/v1/list) are detected separately; a machine with neither makes no request. A Hailo server found at ollama.url is relabelled hailo at start-up, with one notice and an alias migration. /model list and the pickers show Hailo models beside Ollama’s and the catalog the HAT can run but has not pulled, marked not pulled; /model pull [--yes] NAME pulls from it after a confirmation. Harvey never starts hailo-ollama; it prints the systemctl --user command when the card’s server is stopped. A capability table by engine stops harvey sending tools (HTTP 500) or asking for embeddings, tokenize and ps where Hailo has none, and a failed embedding says that Hailo cannot embed and names Ollama.
• Per-model prompt limits. hailo-ollama fails with HTTP 500 once a whole prompt passes about 3,000 characters. /model limit sets a model’s limit (seeded at 700 tokens for Hailo’s llama3.2). With a limit known, the system prompt is fitted to 60% of it (skills catalog first, then the tail of HARVEY.md, then a compact preamble), the oldest whole turns are dropped before each send, and a turn over the limit is refused before it is sent. A turn that fails like an oversize prompt is retried on 70%, 45% and 25% of its size and the first success is remembered as the model’s limit. The tool-loop path does not retry yet.
• harvey completion bash|powershell [-install] writes shell completion for harvey’s verbs, options and help topics, or installs it where the shell loads it. -install never overwrites a file harvey did not write (exit 73). It is separate from the in-session /command completion.
• make test-models runs live model tests (TestLive_*) against whatever model systems the machine has. A missing system skips its tests with the reason and how to start it.
• make generate runs cmt over codemeta.json. make build no longer regenerates version.go, CITATION.cff or about.md; make release runs generate first.

Changed

• The model cache keys capability rows and learned prompt limits by engine and model, so the same name on two engines no longer shares a row. Ollama keeps the bare name; other engines are prefixed. llamafile and llama.cpp rows stored under a bare name are probed once more.
• /memory mine reads a session too large for the model in parts (sized to the model’s prompt limit when known, halved when a request is cut off), merges and de-duplicates the proposals, and reports a model whose limit is too small to hold the mining instructions as unable to mine. An empty session is recorded as empty and never sent; a truncated or unparseable reply leaves the session unmined and is remembered per session and model so auto-mine does not retry the same model. The review prompt quits at end of input.
• A RAG retrieval failure (an embedder that is down) is reported once per distinct cause for the life of the session; the prompt still goes on without context. It used to fail silently on every prompt.
• github.com/rsdoiel/knowledge bumped from v0.0.15 to v0.0.20. Harvey’s own code does not depend on the behaviour changes; the library additions are additive.

Fixed

• Switching model mid-session no longer drops the --debug log or leaves the transcript naming the previous model. Ollama (/model use, the picker, an alias, or a step dispatched with @name or [model: name]) was fixed first; the llama.cpp path was the last one left.
• --ollama and -m/--model were overridden by ollama.url and ollama.model in harvey.yaml; see the upgrade notes for the other flags fixed the same way.
• hailo-ollama models were sent a tools array they cannot take: /api/show returns "model_info": "", which made the capability probe fail to decode, so tool mode stayed auto. The probe now tolerates it and a hef model probes as no tool support.
• A write offer took the next input line as the file path: piping /exit created an empty file named exit.
• Test agents no longer scan the machine’s ~/Models, so the suite gives the same result on every machine.

v0.0.18 (2026-10-02)

Changed

• /profile use’s automatic hand-off (WriteHandoff) now writes Markdown per HANDOFF_FORMAT.md (frontmatter and a State Recap section) with a .md extension, instead of Fountain. It stays workspace-tier and does not touch STATE.md: it is a heuristic snapshot with no LLM pass.
• recallKB retrieves through the knowledge library’s RecallByText (knowledge DR-0051) instead of its own matching.
• github.com/rsdoiel/knowledge bumped from v0.0.14 to v0.0.15. This brings DR-0054 (kb import keeps an observation kind outside the vocabulary and warns) and DR-0055 (kb project add on an existing name exits 1). Harvey’s own code does not depend on either behaviour, but a script that shells out to kb project add should read the knowledge v0.0.15 notes.

Fixed

• An unknown slash command is now a usage error (exit class 2) instead of a printed message that let a scripted session exit 0. At a terminal nothing changes except the message, which now reads Error: unknown command: /name. Inside /loop, a mistyped command is reported on every iteration as before, and /loop now returns the first failure when it finishes (see below). First step of reclassifying the command handlers’ errors (the DR-0006 follow-up).

• The file and workspace commands now return classed errors instead of printing a failure and succeeding, so a scripted session exits with the class of the first one that failed (harvey DR-0006 follow-up). /files, /file-tree, /read, /read-dir, /read-pdf, /read-chunks, /attach, /write, /run, /search, /git, /format, /summarize:

  Failure	Exit
  missing or surplus argument, bad flag value, /search pattern that does not compile, unknown option	2
  no workspace; a named file or directory that does not exist; a program /run cannot start	66
  no backend connected (/read-chunks, /summarize); a remote fetch that fails	69
  read, write or exec refused by Harvey’s permission rules or safe mode; a path outside the workspace, through a symbolic link, in agents/, or matching a sensitive-file pattern	77
  a file of the wrong kind: binary or too large for /attach, not a directory for /read-dir, not a PDF, a source file the formatter rejects	65
  nothing found: /search with no matches, /read-dir with no readable files, /write with no reply to write, /git when git itself fails, /format for an extension with no formatter, /summarize with too little history	1

  /read and /format still try every file, then report the first failure. /run does not fail when the command it ran exits non-zero: that output is the point. At a terminal nothing changes except that the message now appears as Error: ....

• The same for the model, session and settings commands: /model (use, list, show, stop, mode, alias), /inspect, /safemode, /record, /rename, /context, /session (and /resume), /workspace and /help TOPIC. A failed @name model switch in a scripted session now counts as the turn’s failure too.

  Failure	Exit
  missing argument, unknown subcommand, unknown mode or help topic, a model name that matches several models	2
  no workspace; no sessions directory; a session file that does not exist	66
  no backend connected (/session replay), /inspect without an Ollama backend, a model that fails to start	69
  a session file that does not parse	65
  settings that could not be saved (/safemode, /model alias) or a recording that could not be renamed	74
  a recording file that could not be created	73
  no such model, alias, or llamafile; no models or sessions to pick from; a backend to stop that is not running; already recording or not recording; /safemode deny of a command not in the list; no model cache for /model mode	1

  /safemode still applies the change in memory when saving fails, and now also exits 74.

• The same for the memory, knowledge-base, RAG, skill and route commands: /memory (and /memory profile), /kb (project, observe, concept, source, retract, cite, show, learn), /rag, /skill, /skill-set and /route.

  Failure	Exit
  missing argument, unknown subcommand, an ID that is not a number, an unrecognised route URL, an unknown route setting or value, --limit that is not a number	2
  a profile document missing on disk	66
  the knowledge base not open, no backend connected for /skill compile, an encoderfile server that cannot be reached, a model that cannot be used for /kb learn draft, a route whose models cannot be listed	69
  settings that could not be saved (/route, /rag on|off, /memory profile on|off)	74
  a skill-set that could not be created	73
  no such memory, source, observation, route, skill, store or project; nothing to pick from; no current project; /memory recall or /rag query with no results; /rag ingest with no ingestable files; a skill-set that already exists; no skill-set loaded to unload	1

  /kb cite tries every ID, then reports the first failure. /rag ingest does the same across all its sources, local and remote (http://, https://, s3://, sftp://, scp://): a list that cannot be read, a download that fails, an unsupported scheme, or a file the embedding server cannot embed does not stop the rest, and the first failure decides the exit status. Download, list and embedding failures are 69 unless the underlying error already has a class; an unsupported scheme is 2; a temp file that cannot be created is 73; no RAG store (or no active one) for /rag ingest or /rag query is 1.

Fixed (continued)

• /workspace init FROM_PATH never imported anything: the handler read its subcommand from the second argument instead of the first, so through the real dispatcher the path was taken as an unknown subcommand, and /workspace bogus quietly showed the status. The old tests passed a stray "workspace" as the first argument, which hid it.

• /learn found no hand-offs after the workspace moved them from agents/hand-off/ to agents/projects/<project>/hand-off/ (workspace DR-0004). learnCandidates now also scans each project’s hand-off/ directory and accepts .md hand-offs (workspace DR-0005). Session recordings are still Fountain only.

Project layout

• Design briefs, plans, decision records and notes moved out of this repository into ~/Laboratory/agents/projects/harvey/{design,plans,decisions,notes}/ (harvey DR-0008, which supersedes DR-0002). DECISIONS.md is frozen at 2026-09-23 under notes/; new decisions are records. Code comments that cite DECISIONS.md by date are accurate history and were left as they were. developer_guide.md links point at the new locations.

v0.0.17 (2026-09-27)

Breaking (exit codes)

Both harvey and assay now use the workspace exit-code convention (workspace DR-0003; harvey DR-0005) instead of exiting only 0 or 1. A script or an embedding tool that reads either binary’s exit status should check these against the new table (0 ok, 1 negative, 2 usage, 65 data, 66 no_input, 69 unavailable, 70 internal, 73 cant_create, 74 io, 75 temp_fail, 77 no_permission, 78 config — see harvey/CLAUDE.md, “Exit codes”). Five cases that used to exit 0 now fail:

Every other case that used to exit 1 now exits the more specific code for its cause (2 usage, 65 data, 66 no_input, 69 unavailable, 73 cant_create, 77 no_permission — see exit-codes-survey.md’s baseline table and harvey/decisions/0005-*.md for the full list and the reasoning). harvey --help/-h, --version, -l, help, and assay --help/--version are unchanged, at 0. An error nothing classified is 70, never 1 — a 70 anywhere is a bug, not a documented class. Verified old-versus-new against the v0.0.16 baseline with scripts/compare-exit-codes.py: every changed code above and none unexpected, no exit 70, no hang.

Other

• An interactive harvey session still warns and continues on a malformed harvey.yaml or an unreachable backend, so nothing changes at a terminal — only non-interactive runs (stdin not a terminal, or --replay without --replay-continue) now fail fast.
• A surplus positional argument to harvey is now a usage error (2), previously reported as “Unknown flag”.
• assay finishes its run and writes the report before exiting with the class of the first failure; a prompt’s failing automatic checks are still a result, not a tool failure, and still exit 0.

Added (harvey DR-0006)

• -json/--json on harvey and assay: a failing startup/flag-parsing error (or, for harvey, a failed non-interactive session — see below) prints as {"error","class","code"} JSON on stderr instead of text, matching kb’s shape. Recognised wherever it sits on the command line, so it still applies when a later flag — or the flag package’s own parse, for assay — is what fails.
• A non-interactive harvey session (piped stdin, or --replay without --replay-continue) now exits with the class of the first slash command or chat turn that failed, instead of always 0. An interactive session at a terminal is unaffected. /plan’s two previously-unclassified errors (“no backend connected”, “no plan found”) now carry a real class (69, 1); most other command families do not yet and still exit 0 or 70 on a failure — a named follow-up, not done here.
• /rag setup’s embedding-model auto-pick now confirms its keyword-based guess with one live /api/embed call before committing to it, falling back to another candidate (or warning) when the guess turns out to be wrong. ThoroughProbeModel gets its first caller.
• github.com/rsdoiel/termlib bumped to the v0.0.9+354195d pseudo-version (no v0.0.10 tag exists yet).

v0.0.16 (2026-09-24)

New features

• Full agentic-memory tool suite: retrieve_memory, add_memory, update_memory, delete_memory, filter_context, summary_context builtin tools, plus proactive STM-budget warnings
• Unified /model command: /llamafile and /llamacpp merged into one backend-agnostic facade; @mention switches the active model while preserving history; unregistered .llamafile/.gguf models are now found via disk scan
• knowledge module extraction: knowledge-base code split into its own module (github.com/rsdoiel/knowledge), now consumed at the released v0.0.13; recallKB tries concept-tag matching (MatchConceptNames/ RecallByConceptNames) before falling back to substring search, and now surfaces decision records and reviewed document summaries, not just observations
• Cross-machine knowledge.db sync: UUID-based merge tool (bin/kbmerge), legacy experiments→projects migration
• Retraction-checking for cited sources in the knowledge base
• /read-chunks: explicit chunked document analysis, independent of context-overflow triggers
• Knowledge learning mode: /kb learn ingest|draft|review|concepts turns recorded sessions and hand-off notes into searchable, human-reviewed knowledge. A model drafts summaries (@model, else learn_model in agents/harvey.yaml, else the active model); nothing is trusted or searchable until you accept it; $EDITOR edits are recorded as human. /kb learn concepts suggests new concepts, previews each affected document as a diff of changed lines, and writes only after a yes, through the permissions table. Fountain sessions and hand-offs are never rewritten

Bug fixes

• Chunk-prompt guard: two bugs prevented the chunking UX from ever triggering on models with an unknown context limit — fixed and live-verified against Gemma-4-E4B
• Llamafile GPULayers now defaults to 0 (CPU-only) instead of 99, fixing an apparent multi-hour “hang” on Raspberry Pi hardware with no GPU backend
• pickBackend startup picker now lists .gguf/llama.cpp models, not just llamafiles and Ollama
• kb and man added to the safe-mode default command allowlist
• File-write confirmations no longer treat end of input (Ctrl-D, a closed pipe) as “yes”, and tagged code-block writes now honour the permissions: table (previously only the untagged fallback did)
• /kb search no longer fails on hyphenated terms such as map-reduce (picked up from knowledge v0.0.13)

Internal

• R0–R8 refactor series: shared BudgetTracker, resolveDispatchTarget, runBoundedTurn/foldBackTurn helpers; unified sensor/status reporting; go vet and gofmt Check() wired in as sensors
• Removed a stale local replace github.com/rsdoiel/termlib => ../termlib — was silently masking a broken build for anyone without a local ../termlib checkout; now consumes the tagged v0.0.9

v0.0.15 (2026-06-26)

New features

• /model mode [MODEL] {structured|prose|inject|none}: set or display the tool-execution strategy for a model. The mode is persisted in the model cache (agents/model_cache.db) and survives sessions and re-probes. Modes: structured (OpenAI tool_calls), prose (JSON-fence fallback), inject (pre-inject file content, no tool calls), none (plain text).
• File-reference injection (option 1): when a model’s tool mode is inject or when the cache shows it does not reliably call tools, Harvey now automatically scans the user’s prompt for path-like tokens, reads any matching workspace files (≤ 64 KB, text extensions only), and prepends their content as ### File: blocks before sending the prompt. Models that ignore the tools schema can now work with file content without the user manually pasting it.
• Cannot-read retry (option 2): if a model responds with a phrase indicating it cannot access a file (e.g. “I cannot directly read files on your system”), Harvey retries the turn once with the referenced file content pre-loaded via file-reference injection. The retry uses RunToolLoop when the model is in structured mode so tool calls in the retry response are correctly dispatched.
• ModelCapability.ToolMode field and ToolMode* constants added to the model cache. The tool_mode column is added to existing databases automatically on first open (idempotent ALTER TABLE).

Bug fixes

• FastProbeModel no longer sets ToolMode = ToolModeStructured when probing a tool-capable model. Re-probing now preserves any mode the user set via /model mode. Tool dispatch continues to work via the existing SupportsTools == CapYes fallback in toolsReliable().
• Option-2 retry: toolCallRecords is now cleared before the history rollback, preventing rolled-back tool calls from appearing in the Fountain session transcript paired with the retry response.
• Option-2 retry: noToolCalls is now computed correctly after an option-2 retry. A hadToolCalls flag is captured before the rollback so that autoExecuteReply does not fire on the retry response when the first pass executed tool calls.
• Option-2 retry: when useStructuredTools is true, the retry now calls RunToolLoop instead of Client.Chat directly, so structured tool calls in the retry response are correctly dispatched and recorded.
• cantReadPhrases entry "please provide the file" tightened to "please provide the file content" to avoid spurious retries when a model asks for a file path rather than refusing file access.
• Exported ToolMode* constants now have /** … */ block-doc comments per the project documentation convention.

v0.0.14 (2026-06-21)

New features

• Llamafile is now the primary language model backend: startup sequence shows registered llamafiles first, auto-selects on preferred model match, then falls back to Ollama models
• Explicit connection feedback: “Connecting to NAME (llamafile)… ✓” shown in terminal on backend selection and after startAndUseLlamafile
• Stale server adoption: Harvey detects a running llamafile server it did not start, probes its model via /v1/models, warns on model mismatch, and adopts it rather than refusing to start
• /llamafile show [NAME]: displays name, path, file size, and configured context length for a registered model
• /rag show [NAME]: displays store name, database path, embedding model, chunk count, and model map
• Remote RAG ingest extended: sftp://, scp://, http://, and https:// URIs now supported alongside s3://
• /read auto-detects .pdf files and extracts text via poppler (pdfinfo + pdftotext), consistent with the read_file built-in tool
• /status backend tag and token-count estimate now work for llamafile backends (was Ollama-only)
• Pipeline context-utilization display now works for llamafile via character estimate (was Ollama-only)
• Context utilization hint [ctx: N%] added to spinner label when estimated token usage reaches 50% of the model’s context window
• Routing feedback in spinner: shows @route · model during routed turns when routing is active
• Model provenance recorded in Fountain session header: Model: field now stores NAME (backend) (e.g., QWEN-CODING (llamafile)) for session replay and audit
• Health check on --resume/--continue: session model is extracted before backend selection; a mismatch warning is shown when the resumed model differs from the active backend
• @mention dispatch: routing is tried first when routing is enabled and a route exists; falls back to attemptModelSwitch for local model switching; case-insensitive for both llamafile names and model aliases
• Help system: all 41 help-text constants documented with block comments and reordered into 11 logical groups; ModelHelpText dispatch bug fixed (topic was unreachable via /help model); harvey-model.7.md man page added

Documentation

• Complete rewrite of overview.md with “Natural Language Programming for Scholarly Work” framing
• getting_started.md rewritten llamafile-first with correct Mozilla AI model download URL and startup sequence
• user_manual.md restructured: llamafile-first Model Management section, corrected links, expanded man page index
• CONFIGURATION.md fully updated: all v0.0.14 configuration fields documented (model_aliases, llamafile with context_length, tools, memory with rolling_summary and knowledge_base, rag nesting, syntax_highlight, auto_format); SFTP, HTTP, and AWS environment variables added; complete annotated harvey.yaml example

v0.0.13 (2026-06-19)

New features

• /profile command: top-level alias for /memory profile <list|show|edit|use|rename> — profile management without the /memory prefix
• --resume flag: resumes the most recent session in agents/sessions/ automatically at startup (no path needed); prints a notice if no sessions are found
• Spinner live status: tool calls now show a transient “Calling tool…” status line on the second spinner line while waiting for results; StatusReporter interface added to tool_executor.go so the spinner can be wired to the executor
• assay --llamafile PATH: evaluate a llamafile binary directly; assay starts and stops the process automatically and derives the model name from the binary path
• Added web-developer workspace profile template (templates/profiles/web-developer.fountain)
• HARVEY.md: documented native file-reading capabilities — PDF (via poppler) and image (vision routes) — so models know to call read_file directly without asking the user to convert files

Bug fixes

• S3 remote: improved not-found detection for missing keys and missing buckets; path-style access now correctly enabled for non-AWS S3-compatible endpoints (MinIO, Cloudflare R2)
• MostRecentSession helper added to sessions_files.go; --resume reliably selects the newest .spmd/.fountain file by delegating to ListSessionFiles (already sorted newest-first)
• Llamafile: scanLlamafileModels now discovers binaries by directory scan, fixing cases where auto-discovery missed models not registered via /llamafile add

v0.0.12 (2026-06-17)

New features

• Memory enrichment: added kind field to memory documents classifying why knowledge matters (pitfall/workaround/recommendation/pattern)
• Memory enrichment: added action field — the imperative step a future agent should take; included in embedding text for better semantic retrieval
• Memory enrichment: added confidence field (default 0.5); retrieval scores are weighted multiplicatively (score = cosine × confidence)
• /memory flag <id>: new command reduces confidence by 0.1 per call; auto-archives when confidence falls to or below 0.2
• /memory list: new --kind filter; output now shows kind and confidence columns alongside type
• Miner prompt updated to elicit kind and action for each extracted memory; all five memory types now listed
• WriteDigest(): MemoryStore auto-writes agents/memories/DIGEST.md on every Save, Archive, and MineAuto — plain Markdown readable by any LLM without a SQLite client
• agents/skills/harvey-memory/SKILL.md: new cross-agent skill teaching Vibe and Claude Code when and how to use the memory digest
• Memories database lazily migrated: existing memories.db files gain kind, action, confidence columns on first open; FTS5 table rebuilt with new columns
• Added create_dir built-in tool so models can create directories without run_command mkdir
• Added /safe and /safe_mode as aliases for /safemode
• Unknown slash commands now highlighted in yellow
• Tool result compaction: prior tool-call rounds are compacted in RunToolLoop before each new LLM turn, keeping context bounded during multi-step tasks
• /plan command: generate a GFM checklist plan, execute each step with fresh bounded context, track progress in agents/plan.md
• multi-file skill: auto-detects multi-file creation requests and generates a plan via the compiled script path
• Skill dispatch: HARVEY_API_BASE env var added to compiled script environment
• Skill dispatch: LLM-fallback skills now trigger an LLM response turn instead of silently continuing
• Plan execution: steps with blocked or failed tool calls are no longer auto-marked complete

Bug fixes

• Llamafile: fixed exec format error on macOS (APE binaries now launched via /bin/sh)
• Llamafile: added --server flag for headless mode (llamafile v0.10.3 API change)
• Llamafile: added -ngl GPU layer offload support with gpu_layers config option (default 99, maximises Metal/CUDA)
• Llamafile: startup_timeout config option (default 120s); fast-fail on process exit with stderr surfaced in error
• Llamafile: debug log now wired to new client after /llamafile use model switch
• Skill trigger regex: fixed /pattern/flags format (trailing flag suffix no longer breaks regex mode)
• Skill dispatch: compilation failure now falls back to LLM context-injection path instead of erroring out

v0.0.11 (2026-06-11)

New features

• Added scholarly identifier extraction and normalization for 14 identifier types (DOI, ORCID, ROR, RAiD, ArXiv, FundRef, ISBN, ISSN, ISNI, PMID, PMCID, VIAF, SNAC, LCNAF) via scholarly_identifiers.go and github.com/caltechlibrary/metadatatools
• Added scholarly-aware PDF ingest: papers are chunked by section (abstract, introduction, methods, results, discussion, conclusion, references) and tagged with the document’s own identifiers and any cited works’ identifiers
• Extended the knowledge base schema so observations can record a source DOI and concepts can represent scholarly entities (people, papers, institutions, funders) via an identifier type/value pair
• Workspace onboarding now scans codemeta.json/CITATION.cff for identifiers (e.g. author ORCID iDs, release DOIs) and records them in the project’s project_fact memory metadata

v0.0.10 (2026-06-09)

New features

• Added multi-language code-aware chunking for RAG ingestion (C, C++, Pascal, Oberon, Lisp, Basic)
• Added documentation extraction: comment and docstring association with symbols for C, C++, Pascal, Oberon, Lisp, Basic
• Added ANSI syntax highlighting of code blocks in LLM responses (13 languages: C, C++, Pascal, Oberon, Lisp, Basic, Go, Python, JavaScript, TypeScript, Rust, Shell, SQL); configurable via syntax_highlight in harvey.yaml
• Added automatic code formatting on write_file: built-in formatters for Pascal, Oberon, Basic; external pipe-mode formatters for Go (gofmt), C/C++ (clang-format), Python (black), Rust (rustfmt), JavaScript/TypeScript (prettier); configurable via auto_format in harvey.yaml
• Added /format FILE [FILE...] command to manually format workspace source files in-place

v0.0.9 (2026-06-09)

Added /loop command

v0.0.8 (2026-06-05)

Initial profile template system

• Added /profile use command for workspace profile management
• Added help guides and /status profile display

v0.0.7 (2026-06-02)

New features

assay evaluation harness — a new bin/assay tool for running a prompt corpus against one or more Ollama models and producing a Markdown report plus JSON results for human review and automated checking.

• Corpus defined in YAML (agents/assay/corpus.yaml); each prompt specifies category, language, automated checks (contains, not_contains, compiles, go_vet), and human-review questions
• --rag-db PATH — opens a RAG store and injects retrieved context before each prompt, enabling RAG-assisted evaluation
• --rag-compare — runs each prompt twice (base + RAG) and writes a per-check delta table alongside the main report
• --category NAME — run only one category of prompts (e.g. go-crosswalk)
• Summary table (model × prompts × auto-pass rate × average tok/s) at the top of every report

File attachment commands

• /attach FILE — route-aware file injection: native image (JPEG, PNG, GIF, WebP) for vision-capable routes, text extraction for PDFs, plain-text for source files; not restricted to the workspace
• /read-pdf FILE [PAGES] — extract text from a PDF using poppler utilities (pdfinfo, pdftotext, pdfimages) and inject into context; accepts a page range (e.g. 40-55); cap of 20 pages per call; diagram-only pages are flagged; not restricted to the workspace

Knowledge discoverability

• /hint — on-demand improvement suggestions: flags unmined sessions, empty or disabled RAG stores, and empty knowledge base
• /recall QUERY — alias for /memory recall; searches all three knowledge silos (RAG, memory store, knowledge base) in one call
• /help learn — new unified help topic explaining the three-silo architecture and the single decision rule for where to put each type of content
• Session-start memory digest: on REPL startup Harvey prints dim actionable hints when sessions are unmined, the active RAG store is empty, or RAG is off with chunks present
• Enhanced /status — now includes a Memory/RAG summary block (active memories injected, unmined session count, active RAG store, chunk count, RAG on/off)

Persistent command history — the REPL input history is saved to agents/harvey_history on exit (capped at 1000 entries, most recent kept) and reloaded on the next startup; history is per-workspace to avoid leaking commands and paths between projects.

Bug fixes

• Removed unreachable dead code (cmdRunCtx, extractRunSuggestions) that was missing the safe-mode allowlist check and would have bypassed it had it ever been wired into an execution path
• Help text overview (/help) was missing entries for /attach, /read-pdf, /hint, /memory, /recall, and /pipeline; all are now listed under their appropriate sections
• harvey --help attach, --help read-pdf, and --help learn now work correctly; topic was silently falling through to the unknown-topic error

v0.0.6 (2026-05-30)

New features

Remote protocol integration — Harvey can now read files from remote storage over s3://, http://, https://, sftp://, and scp:// URIs anywhere a local path was accepted (/read, /attach, RAG ingest).

• RemoteReader interface with URI-scheme factory (NewRemoteReader)
• S3-compatible backend via MinIO Go client (AWS S3, MinIO, Cloudflare R2)
• HTTP/HTTPS backend (stdlib net/http); HTTPS→HTTP downgrade protection
• SFTP/SCP backend (golang.org/x/crypto/ssh); strict host-key fingerprint verification required — anonymous connections are refused
• S3 RAG ingest: /rag add s3://bucket/prefix/ ingests objects as RAG docs
• Tab completion extended to remote URI prefixes

Unified memory system

• MemoryConfig is now the single canonical owner of all RAG and knowledge-base configuration; memory.rag: and memory.knowledge_base: replace top-level rag: in harvey.yaml (old format still loaded for backward compatibility)
• Rolling summary: when conversation history exceeds memory.budget_pct (default 25 %) of the model context window, Harvey compresses all but the last six turns into a ~150-token summary with a warning
• Hybrid retrieval: FTS5 lexical fast-path combined with cosine similarity (when an embedder is active); Recent() fallback replaced by relevance ranking
• Session stats: memory_stats table in memories.db tracks token budget, injected tokens, compression events, and average tokens/sec per session
• /memory status now shows budget utilisation and, after ten sessions, prints evidence-based advice to increase or reduce budget_pct
• /memory recall QUERY — on-demand semantic search across all memory types
• Workspace-profile onboarding: on first run Harvey prompts four questions and writes a workspace_profile memory document; no longer asks for workspace name (derived from the directory)

Pipeline support

• /pipeline command and implementation for multi-step prompt pipelines
• Pipeline steps can reference workspace files and inject their output into subsequent steps

Security fixes

• SFTP — OOM prevention: sftpReadPacket now rejects packets larger than 4 MB; sftpReadStr rejects strings larger than 256 KB. A malicious or misconfigured SFTP server could previously trigger a ~4 GB allocation.
• HTTP — response body limit: Get caps the response body at 256 MB via io.LimitReader; responses that hit the cap return an error instead of exhausting process memory.
• HTTP — client timeout: http.Client.Timeout is now 5 minutes, preventing indefinite hangs when a server stalls after accepting the connection.
• Credential filtering: filterCommandEnvironment extended to strip AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, AWS_SECURITY_TOKEN, MINIO_ACCESS_KEY, MINIO_SECRET_KEY, SFTP_PASSWORD, SFTP_KEY_PATH, HTTP_BEARER_TOKEN, HTTP_BASIC_PASSWORD from child-process environments.

Bug fixes

• Fixed tools-loop processing problem where the model’s tool-call responses were not advancing the conversation correctly
• Fixed write_file tool failure: malformed JSON in model responses no longer silently drops the write; error is surfaced to the user
• Fixed missing help documentation for several commands added in v0.0.5

v0.0.5 (2026-05-17)

• Initial public release
• Terminal REPL with Ollama backend
• Tool calling: read_file, write_file, list_files, run_command, git
• RAG store support (SQLite-backed vector search)
• Knowledge-base integration (/kb commands)
• Session recording to Fountain screenplay format (.spmd)
• Safe mode with command allowlist
• Multi-model routing (/route, @name dispatch)
• Skill dispatch system
• Cross-platform installers (Linux x86-64/aarch64/armv7l, macOS, Windows)